def __init__(self, name="car_racing", renders=False, record_data=False, is_discrete=True, state_dim=-1,
learn_states=False, save_path='srl_zoo/data/', srl_model="raw_pixels", env_rank=0, srl_pipe=None,
lookahead=5, **_):
"""
Gym wrapper for Racing car environment
WARNING: to be compatible with kuka scripts, additional keyword arguments are discarded
:param name: (str) name of the folder where recorded data will be stored
:param renders: (bool) Whether to display the GUI or not
:param record_data: (bool) Set to true, record frames with the rewards.
:param is_discrete: (bool) Whether to use discrete or continuous actions
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param save_path: (str) location where the saved data should go
:param srl_model: (str) The SRL_model used
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param lookahead: (int) How many segments ahead of the current position of the track should the target be
"""
SRLGymEnv.__init__(self, srl_model=srl_model, relative_pos=RELATIVE_POS, env_rank=env_rank, srl_pipe=srl_pipe)
GymCarRacing.__init__(self)
self._renders = renders
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
self._is_discrete = is_discrete
self.lookahead = lookahead
self.relative_pos = RELATIVE_POS
self._env_step_counter = 0
self._observation = None
self.saver = None
if record_data:
self.saver = EpisodeSaver(name, None, state_dim, globals_=getGlobals(), relative_pos=RELATIVE_POS,
learn_states=learn_states, path=save_path)
# Accelerate, brake, stear left, stear right
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1, high=1, shape=(3,), dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0, high=255, shape=(self._height, self._width, 3), dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(self.state_dim,), dtype=np.float32)
def __init__(self,
name="RLBenchDS-unset",
wrapped_env=None,
max_steps_per_epoch=100,
path=None,
**_kwargs):
#Passed in in dataset_generator.py kwargs.
self.env = wrapped_env
self.max_steps_per_epoch = max_steps_per_epoch
print("Recording data...")
try:
env_name = self.env.task.get_name()
except:
env_name = self.env.unwrapped.spec.id
print(env_name)
self.saver = EpisodeSaver(name, env_name=env_name, path=path)
self.action_space = self.env.action_space
def __init__(self, renders=False, is_discrete=True,
name="labyrinth", max_distance=1.6, shape_reward=False, record_data=False, srl_model="raw_pixels",
random_target=False, state_dim=-1, verbose=False, save_path='srl_zoo/data/',
env_rank=0, srl_pipe=None, img_shape=None, **_):
super(LabyrinthEnv1, self).__init__(srl_model=srl_model,
relative_pos=False,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._observation = []
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
# self._shape_reward = shape_reward
self._random_target = random_target
self.state_dim = state_dim
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(2, dtype=int) # index on the self.map array
self.target_pos = [] # list of indexes (target positions) on the self.map array
self.previous_robot_pos = None
self.has_bumped = False # Used for handling collisions
self.count_collected_tresors = 0
self._env_step_counter = 0
self.srl_model = srl_model
self.maze_size = MAZE_SIZE
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
if record_data:
self.saver = EpisodeSaver(name, max_distance, state_dim, globals_=getGlobals(), path=save_path)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0, high=255, shape=(self._height, self._width, 3), dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(self.state_dim,), dtype=np.float32)
def __init__(self,
renders=False,
is_discrete=True,
name="mobile_robot",
max_distance=1.6,
shape_reward=False,
record_data=False,
srl_model="raw_pixels",
random_target=False,
state_dim=-1,
verbose=False,
save_path='srl_zoo/data/',
env_rank=0,
srl_pipe=None,
img_shape=None,
**_):
super(MobileRobotX, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._is_discrete = is_discrete
self.state_dim = state_dim
self.relative_pos = RELATIVE_POS
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(2)
self.target_pos = np.zeros(2)
# Boundaries of the square env
self._min_x, self._max_x = 0, 4
self._min_y, self._max_y = 0, 4
self.collision_margin = 0.1
self._observation = []
self._env_step_counter = 0
self.has_bumped = False # Used for handling collisions
self.imgs_folder = "mobilerobot" ## 'mobilerobot', 'omnirobot'
self.robot_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/robot_224.png".
format(self.imgs_folder))
self.target_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/target_224.png".
format(self.imgs_folder))
self.background_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/background_224.png".
format(self.imgs_folder))
robot_img_shape = self.robot_img.shape
self.robot_img = cv2.resize(
self.robot_img, (int(robot_img_shape[1] * self._width / 224),
int(robot_img_shape[0] * self._height / 224)))
target_img_shape = self.target_img.shape
self.target_img = cv2.resize(
self.target_img, (int(target_img_shape[1] * self._width / 224),
int(target_img_shape[0] * self._height / 224)))
self.background_img = cv2.resize(self.background_img,
(self._width, self._height))
# cv2.imshow("robot", self.robot_img)
# k = cv2.waitKey(0)
# if k == ord("q"):
# raise KeyboardInterrupt
# self.walls = None
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name,
max_distance,
state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=False,
path=save_path)
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1,
high=1,
shape=(2, ),
dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
class MobileRobotX(SRLGymEnv):
"""
"""
def __init__(self,
renders=False,
is_discrete=True,
name="mobile_robot",
max_distance=1.6,
shape_reward=False,
record_data=False,
srl_model="raw_pixels",
random_target=False,
state_dim=-1,
verbose=False,
save_path='srl_zoo/data/',
env_rank=0,
srl_pipe=None,
img_shape=None,
**_):
super(MobileRobotX, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._is_discrete = is_discrete
self.state_dim = state_dim
self.relative_pos = RELATIVE_POS
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(2)
self.target_pos = np.zeros(2)
# Boundaries of the square env
self._min_x, self._max_x = 0, 4
self._min_y, self._max_y = 0, 4
self.collision_margin = 0.1
self._observation = []
self._env_step_counter = 0
self.has_bumped = False # Used for handling collisions
self.imgs_folder = "mobilerobot" ## 'mobilerobot', 'omnirobot'
self.robot_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/robot_224.png".
format(self.imgs_folder))
self.target_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/target_224.png".
format(self.imgs_folder))
self.background_img = cv2.imread(
"./environments/mobile_robot_extreme/images/{}/background_224.png".
format(self.imgs_folder))
robot_img_shape = self.robot_img.shape
self.robot_img = cv2.resize(
self.robot_img, (int(robot_img_shape[1] * self._width / 224),
int(robot_img_shape[0] * self._height / 224)))
target_img_shape = self.target_img.shape
self.target_img = cv2.resize(
self.target_img, (int(target_img_shape[1] * self._width / 224),
int(target_img_shape[0] * self._height / 224)))
self.background_img = cv2.resize(self.background_img,
(self._width, self._height))
# cv2.imshow("robot", self.robot_img)
# k = cv2.waitKey(0)
# if k == ord("q"):
# raise KeyboardInterrupt
# self.walls = None
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name,
max_distance,
state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=False,
path=save_path)
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1,
high=1,
shape=(2, ),
dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
def getTargetPos(self):
# Return only the [x, y] coordinates
return self.target_pos
# @staticmethod
def getGroundTruthDim(self):
"""
The new convention for the ground truth (GT): GT should include target position under random target
setting.
"""
# HACK : Monkey-Patch, shit solution to solve problem.
if not self._random_target:
return 2
else:
return 4
def getRobotPos(self):
# Return only the [x, y] coordinates
return self.robot_pos
def getGroundTruth(self):
"""
The new convention for the ground truth (GT): GT should include target position under random target
setting. A better solution would be "change all the environment files, especially srl_env.py" !!!
"""
# HACK: Monkey-Patch, shit solution to solve problem.
robot_pos = self.getRobotPos()
if self._random_target:
if self.relative_pos:
# HACK here! Change srl_env.py and all the other envs in the future !!! TODO TODO TODO
return robot_pos
else:
# check 'envs.observation_space' in rl_baselines/base_classes.py (before self.model.learn) !!!
target_pos = self.getTargetPos()
return np.concatenate([robot_pos, target_pos], axis=0)
else:
if self.relative_pos:
# HACK here! Change srl_env.py and all the other envs in the future !!! TODO TODO TODO
return robot_pos
else:
return robot_pos
def reset(self):
self._env_step_counter = 0
self.has_bumped = False
# Init the robot randomly
x_start = self._max_x / 2 + self.np_random.uniform(
-self._max_x / 3, self._max_x / 3)
y_start = self._max_y / 2 + self.np_random.uniform(
-self._max_y / 3, self._max_y / 3)
self.robot_pos = np.array([x_start, y_start])
# Initialize target position
x_pos = 0.8 * self._max_x
y_pos = 0.7 * self._max_y
if self._random_target:
margin = self.collision_margin * self._max_x + 0.08 ## 0.4 + 0.05
x_pos = self.np_random.uniform(self._min_x + margin,
self._max_x - margin)
y_pos = self.np_random.uniform(self._min_y + margin,
self._max_y - margin)
self.target_pos = np.array([x_pos, y_pos])
# self.walls = [wall_left, wall_bottom, wall_right, wall_top]
self._observation = self.getObservation()
if self.saver is not None:
self.saver.reset(self._observation, self.getTargetPos(),
self.getRobotPos())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation)
return self._observation
def getObservation(self):
"""
:return: (numpy array)
"""
self._observation = self.render("rgb_array")
return self._observation
def step(self, action):
# True if it has bumped against a wall
self._env_step_counter += 1
self.has_bumped = False
if self._is_discrete:
dv = DELTA_POS
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD)
dx = [-dv, dv, 0, 0][action]
dy = [0, 0, -dv, dv][action]
real_action = np.array([dx, dy])
else:
dv = DELTA_POS
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD)
# scale action amplitude between -dv and dv
real_action = np.maximum(np.minimum(action, 1), -1) * dv
previous_pos = self.robot_pos.copy()
self.robot_pos += real_action
# Handle collisions
for i, (limit, robot_dim) in enumerate(
zip([self._max_y, self._max_x], [ROBOT_WIDTH, ROBOT_LENGTH])):
margin = self.collision_margin * limit + robot_dim / 2 #
margin += 0.02
# If it has bumped against a wall, stay at the previous position
if self.robot_pos[i] <= margin or self.robot_pos[
i] >= limit - margin:
self.has_bumped = True
self.robot_pos = previous_pos
break
# Update mobile robot position on image
self._observation = self.getObservation()
# calculate the reward of this step
distance = np.linalg.norm(self.getTargetPos() - self.robot_pos, 2)
reward = 0
if distance <= REWARD_DIST_THRESHOLD:
reward = 1
# Negative reward when it bumps into a wall
if self.has_bumped:
reward = -1
if self._shape_reward:
reward = -distance
done = self._env_step_counter > self.max_steps
if self.saver is not None:
self.saver.step(self._observation, action, reward, done,
self.getRobotPos())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation), reward, done, {}
return np.array(self._observation), reward, done, {}
def render(self, mode='rgb_array'):
"""
mode is useless
"""
image = self.background_img.copy()
# self.robot_pos = np.array([1.2, 3.6])
# ## get robot position on image (pixel position)
# robot_img_pos = np.array([(self.robot_pos[1]/self._max_y)*self._height, (self.robot_pos[0]/self._max_x) * self._width], dtype=int)
# h_st = robot_img_pos[0]-self.robot_img.shape[0]//2 #, robot_img_pos[0]+self.robot_img.shape[0]//2
# h_ed = self.robot_img.shape[0]+h_st
# w_st = robot_img_pos[1]-self.robot_img.shape[1]//2
# w_ed = self.robot_img.shape[1]+w_st
# # print(self.robot_pos)
# ## HACK
# mask = (np.mean(self.robot_img, axis=-1) < 250)
# # import matplotlib.pyplot as plt
# # plt.figure()
# # plt.imshow(mask.astype(float))
# # plt.show()
# image[h_st:h_ed, w_st:w_ed, :][mask] = self.robot_img[mask]
# self.target_pos = np.array([3.6, 3.6])
# print(self.target_pos)
for obj_pos, obj_img in zip([self.target_pos, self.robot_pos],
[self.target_img, self.robot_img]):
obj_img_pos = np.array([(obj_pos[1] / self._max_y) * self._height,
(obj_pos[0] / self._max_x) * self._width],
dtype=int)
h_st = obj_img_pos[0] - obj_img.shape[0] // 2
h_ed = obj_img.shape[0] + h_st
w_st = obj_img_pos[1] - obj_img.shape[1] // 2
w_ed = obj_img.shape[1] + w_st
mask = (np.mean(obj_img, axis=-1) < 240)
image[h_st:h_ed, w_st:w_ed, :][mask] = obj_img[mask]
if self._renders:
cv2.imshow(
"Keep pressing (any key) to display or 'q'/'Esc' to quit.",
image)
key = cv2.waitKey(0)
if key == ord('q') or key == 27: # 'q' or 'Esc'
cv2.destroyAllWindows()
raise KeyboardInterrupt
return image[..., ::-1]
def interactive(self, show_map=False):
image = self.getObservation()
cv2.imshow("image", image[..., ::-1])
while True:
k = cv2.waitKey(0)
if k == 27 or k == ord("q"): # press 'Esc' or 'q' to quit
break
elif k == ord("2") or k == 84:
# down
action = 3
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("6") or k == 83:
# right
action = 1
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("8") or k == 82:
# up
action = 2
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("4") or k == 81:
# left
action = 0
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
else:
print("You are pressing the key: {}".format(k))
def __init__(self, urdf_root=pybullet_data.getDataPath(), renders=False, is_discrete=True, multi_view=False,
name="kuka_button_gym", max_distance=0.8, action_repeat=1, shape_reward=False, action_joints=False,
record_data=False, random_target=False, force_down=True, state_dim=-1, learn_states=False,
verbose=False, save_path='srl_zoo/data/', env_rank=0, srl_pipe=None, srl_model="raw_pixels", **_):
super(InmoovButtonGymEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._timestep = 1. / 240.
# URDF PATH
self._urdf_root = urdf_root
self._urdf_sjtu = "../../urdf_robot/"
self._action_repeat = action_repeat
self._observation = []
self._env_step_counter = 0
self._renders = renders
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
self._cam_dist = 1.1
self._cam_yaw = 145
self._cam_pitch = -36
self._cam_roll = 0
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._force_down = force_down
self.camera_target_pos = (0.316, -0.2, -0.1)
self._is_discrete = is_discrete
self.terminated = False
self.renderer = p.ER_TINY_RENDERER
self.debug = True
self.n_contacts = 0
self.state_dim = state_dim
self.action_joints = action_joints
self.relative_pos = RELATIVE_POS
self.cuda = th.cuda.is_available()
self.saver = None
self.multi_view = multi_view
self.verbose = verbose
self.max_steps = MAX_STEPS
self.n_steps_outside = 0
self.table_uid = None
self.button_pos = None
self.button_uid = None
self._kuka = None
self._inmoov = None
self.action = None
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name, max_distance, state_dim, globals_=getGlobals(), relative_pos=RELATIVE_POS,
learn_states=learn_states, path=save_path)
if self._renders:
client_id = p.connect(p.SHARED_MEMORY)
if client_id < 0:
p.connect(p.GUI)
p.resetDebugVisualizerCamera(1.3, 180, -41, [0.52, -0.2, -0.33])
self.debug = True
# Debug sliders for moving the camera
self.x_slider = p.addUserDebugParameter("x_slider", -10, 10, self.camera_target_pos[0])
self.y_slider = p.addUserDebugParameter("y_slider", -10, 10, self.camera_target_pos[1])
self.z_slider = p.addUserDebugParameter("z_slider", -10, 10, self.camera_target_pos[2])
self.dist_slider = p.addUserDebugParameter("cam_dist", 0, 10, self._cam_dist)
self.yaw_slider = p.addUserDebugParameter("cam_yaw", -180, 180, self._cam_yaw)
self.pitch_slider = p.addUserDebugParameter("cam_pitch", -180, 180, self._cam_pitch)
else:
p.connect(p.DIRECT)
global CONNECTED_TO_SIMULATOR
CONNECTED_TO_SIMULATOR = True
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
if self.action_joints:
# 7 angles for the arm rotation, from -1 to 1
#TODO:action_dim should be changed
action_dim = 7
self._action_bound = 1
else:
# 3 direction for the arm movement, from -1 to 1
action_dim = 3
self._action_bound = 1
action_high = np.array([self._action_bound] * action_dim)
self.action_space = spaces.Box(-action_high, action_high, dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
class InmoovButtonGymEnv(SRLGymEnv):
"""
Gym wrapper for Kuka environment with a push button
:param urdf_root: (str) Path to pybullet urdf files
:param renders: (bool) Whether to display the GUI or not
:param is_discrete: (bool) Whether to use discrete or continuous actions
:param multi_view :(bool) if TRUE -> returns stacked images of the scene on 6 channels (two cameras)
:param name: (str) name of the folder where recorded data will be stored
:param max_distance: (float) Max distance between end effector and the button (for negative reward)
:param action_repeat: (int) Number of timesteps an action is repeated (here it is equivalent to frameskip)
:param shape_reward: (bool) Set to true, reward = -distance_to_goal
:param action_joints: (bool) Set actions to apply to the joint space
:param record_data: (bool) Set to true, record frames with the rewards.
:param random_target: (bool) Set the button position to a random position on the table
:param force_down: (bool) Set Down as the only vertical action allowed
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param verbose: (bool) Whether to print some debug info
:param save_path: (str) location where the saved data should go
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param srl_model: (str) The SRL_model used
"""
def __init__(self, urdf_root=pybullet_data.getDataPath(), renders=False, is_discrete=True, multi_view=False,
name="kuka_button_gym", max_distance=0.8, action_repeat=1, shape_reward=False, action_joints=False,
record_data=False, random_target=False, force_down=True, state_dim=-1, learn_states=False,
verbose=False, save_path='srl_zoo/data/', env_rank=0, srl_pipe=None, srl_model="raw_pixels", **_):
super(InmoovButtonGymEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._timestep = 1. / 240.
# URDF PATH
self._urdf_root = urdf_root
self._urdf_sjtu = "../../urdf_robot/"
self._action_repeat = action_repeat
self._observation = []
self._env_step_counter = 0
self._renders = renders
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
self._cam_dist = 1.1
self._cam_yaw = 145
self._cam_pitch = -36
self._cam_roll = 0
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._force_down = force_down
self.camera_target_pos = (0.316, -0.2, -0.1)
self._is_discrete = is_discrete
self.terminated = False
self.renderer = p.ER_TINY_RENDERER
self.debug = True
self.n_contacts = 0
self.state_dim = state_dim
self.action_joints = action_joints
self.relative_pos = RELATIVE_POS
self.cuda = th.cuda.is_available()
self.saver = None
self.multi_view = multi_view
self.verbose = verbose
self.max_steps = MAX_STEPS
self.n_steps_outside = 0
self.table_uid = None
self.button_pos = None
self.button_uid = None
self._kuka = None
self._inmoov = None
self.action = None
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name, max_distance, state_dim, globals_=getGlobals(), relative_pos=RELATIVE_POS,
learn_states=learn_states, path=save_path)
if self._renders:
client_id = p.connect(p.SHARED_MEMORY)
if client_id < 0:
p.connect(p.GUI)
p.resetDebugVisualizerCamera(1.3, 180, -41, [0.52, -0.2, -0.33])
self.debug = True
# Debug sliders for moving the camera
self.x_slider = p.addUserDebugParameter("x_slider", -10, 10, self.camera_target_pos[0])
self.y_slider = p.addUserDebugParameter("y_slider", -10, 10, self.camera_target_pos[1])
self.z_slider = p.addUserDebugParameter("z_slider", -10, 10, self.camera_target_pos[2])
self.dist_slider = p.addUserDebugParameter("cam_dist", 0, 10, self._cam_dist)
self.yaw_slider = p.addUserDebugParameter("cam_yaw", -180, 180, self._cam_yaw)
self.pitch_slider = p.addUserDebugParameter("cam_pitch", -180, 180, self._cam_pitch)
else:
p.connect(p.DIRECT)
global CONNECTED_TO_SIMULATOR
CONNECTED_TO_SIMULATOR = True
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
if self.action_joints:
# 7 angles for the arm rotation, from -1 to 1
#TODO:action_dim should be changed
action_dim = 7
self._action_bound = 1
else:
# 3 direction for the arm movement, from -1 to 1
action_dim = 3
self._action_bound = 1
action_high = np.array([self._action_bound] * action_dim)
self.action_space = spaces.Box(-action_high, action_high, dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
#elif self.srl_model == "joints":
# self.state_dim = self.getJointsDim()
#elif self.srl_model == "joints_position":
# self.state_dim = self.getGroundTruthDim() + self.getJointsDim()
#if self.srl_model == "raw_pixels":
# self.observation_space = spaces.Box(low=0, high=255, shape=(self._height, self._width, 3), dtype=np.uint8)
#else:
# self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(self.state_dim,), dtype=np.float32)
def getSRLState(self, observation):
state = []
if self.srl_model in ["ground_truth"]:
if self.relative_pos:
state += list(self.getGroundTruth() - self.getTargetPos())
else:
state += list(self.getGroundTruth())
if len(state) != 0:
return np.array(state)
else:
self.srl_pipe[0].put((self.env_rank, observation))
return self.srl_pipe[1][self.env_rank].get()
'''
if self.srl_model in ["ground_truth", "joints_position"]:
if self.relative_pos:
state += list(self.getGroundTruth() - self.getTargetPos())
else:
state += list(self.getGroundTruth())
if self.srl_model in ["joints", "joints_position"]:
state += list(self._kuka.joint_positions)
if len(state) != 0:
return np.array(state)
else:
self.srl_pipe[0].put((self.env_rank, observation))
return self.srl_pipe[1][self.env_rank].get()'''
def getTargetPos(self):
return self.button_pos
@staticmethod
def getJointsDim():
"""
:return: (int)
"""
return 6
@staticmethod
def getGroundTruthDim():
return 3
def getGroundTruth(self):
return np.array(self.getArmPos())
def getArmPos(self):
"""
:return: ([float]) Position (x, y, z) of inmoov hand
"""
return p.getLinkState(self._inmoov.inmoov_id, self._inmoov.effectorId)[0]
def reset(self):
self.terminated = False
self.n_contacts = 0
self.n_steps_outside = 0
p.resetSimulation()
p.setPhysicsEngineParameter(numSolverIterations=150)
p.setTimeStep(self._timestep)
p.loadURDF(os.path.join(self._urdf_root, "plane.urdf"), [0, 0, -1])
self.table_uid = p.loadURDF(os.path.join(self._urdf_root, "table/table.urdf"), 0.5000000, 0.00000, -.820000,
0.000000, 0.000000, 0.0, 1.0)
# Initialize button position
x_pos = 0.5
y_pos = 0
if self._random_target:
x_pos += 0.15 * self.np_random.uniform(-1, 1)
y_pos += 0.3 * self.np_random.uniform(-1, 1)
self.button_uid = p.loadURDF("/urdf/simple_button.urdf", [x_pos, y_pos, Z_TABLE])
self.button_pos = np.array([x_pos, y_pos, Z_TABLE])
p.setGravity(0, 0, -10)
self._inmoov = inmoov.Inmoov(urdf_path = self._urdf_sjtu)
self._env_step_counter = 0
# Close the gripper and wait for the arm to be in rest position
for _ in range(500):
if self.action_joints:
self._inmoov.applyAction(list(np.array(self._inmoov.joint_positions)[:7]) + [0, 0])
else:
self._inmoov.applyAction([0, 0, 0, 0, 0])
p.stepSimulation()
# Randomize init arm pos: take 5 random actions
for _ in range(N_RANDOM_ACTIONS_AT_INIT):
if self._is_discrete:
action = [0, 0, 0, 0, 0]
sign = 1 if self.np_random.rand() > 0.5 else -1
action_idx = self.np_random.randint(3) # dx, dy or dz
action[action_idx] += sign * DELTA_V
else:
if self.action_joints:
joints = np.array(self._kuka.joint_positions)[:7]
joints += DELTA_THETA * self.np_random.normal(joints.shape)
action = list(joints) + [0, 0]
else:
action = np.zeros(5)
rand_direction = self.np_random.normal((3,))
# L2 normalize, so that the random direction is not too high or too low
rand_direction /= np.linalg.norm(rand_direction, 2)
action[:3] += DELTA_V_CONTINUOUS * rand_direction
self._kuka.applyAction(list(action))
p.stepSimulation()
self._observation = self.getExtendedObservation()
self.button_pos = np.array(p.getLinkState(self.button_uid, BUTTON_LINK_IDX)[0])
self.button_pos[2] += BUTTON_DISTANCE_HEIGHT # Set the target position on the top of the button
if self.saver is not None:
self.saver.reset(self._observation, self.getTargetPos(), self.getGroundTruth())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation)
return np.array(self._observation)
def __del__(self):
if CONNECTED_TO_SIMULATOR:
p.disconnect()
def getExtendedObservation(self):
if getNChannels() > 3:
self.multi_view = True
self._observation = self.render("rgb_array")
return self._observation
def step(self, action):
# if you choose to do nothing
if action is None:
if self.action_joints:
return self.step2(list(np.array(self._kuka.joint_positions)[:7]) + [0, 0])
else:
return self.step2([0, 0, 0, 0, 0])
self.action = action # For saver
if self._is_discrete:
dv = DELTA_V # velocity per physics step.
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD)
dx = [-dv, dv, 0, 0, 0, 0][action]
dy = [0, 0, -dv, dv, 0, 0][action]
if self._force_down:
dz = [0, 0, 0, 0, -dv, -dv][action] # Remove up action
else:
dz = [0, 0, 0, 0, -dv, dv][action]
# da = [0, 0, 0, 0, 0, -0.1, 0.1][action] # end effector angle
finger_angle = 0.0 # Close the gripper
# real_action = [dx, dy, -0.002, da, finger_angle]
real_action = [dx, dy, dz, 0, finger_angle]
print("Action: [dx, dy, dz, 0, finger_angle]: {}".format(real_action))
else:
if self.action_joints:
arm_joints = np.array(self._kuka.joint_positions)[:7]
d_theta = DELTA_THETA
# Add noise to action
d_theta += self.np_random.normal(0.0, scale=NOISE_STD_JOINTS)
# append [0,0] for finger angles
real_action = list(action * d_theta + arm_joints) + [0, 0] # TODO remove up action
else:
dv = DELTA_V_CONTINUOUS
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD_CONTINUOUS)
dx = action[0] * dv
dy = action[1] * dv
if self._force_down:
dz = -abs(action[2] * dv) # Remove up action
else:
dz = action[2] * dv
finger_angle = 0.0 # Close the gripper
real_action = [dx, dy, dz, 0, finger_angle]
if self.verbose:
print(np.array2string(np.array(real_action), precision=2))
return self.step2(real_action)
def step2(self, action):
"""
:param action:([float])
"""
# Apply force to the button
p.setJointMotorControl2(self.button_uid, BUTTON_GLIDER_IDX, controlMode=p.POSITION_CONTROL, targetPosition=0.1)
for i in range(self._action_repeat):
self._kuka.applyAction(action)
p.stepSimulation()
if self._termination():
break
self._env_step_counter += 1
self._observation = self.getExtendedObservation()
if self._renders:
time.sleep(self._timestep)
reward = self._reward()
done = self._termination()
if self.saver is not None:
self.saver.step(self._observation, self.action, reward, done, self.getGroundTruth())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation), reward, done, {}
return np.array(self._observation), reward, done, {}
def render(self, mode='human', close=False):
if mode != "rgb_array":
return np.array([])
camera_target_pos = self.camera_target_pos
if self.debug:
self._cam_dist = p.readUserDebugParameter(self.dist_slider)
self._cam_yaw = p.readUserDebugParameter(self.yaw_slider)
self._cam_pitch = p.readUserDebugParameter(self.pitch_slider)
x = p.readUserDebugParameter(self.x_slider)
y = p.readUserDebugParameter(self.y_slider)
z = p.readUserDebugParameter(self.z_slider)
camera_target_pos = (x, y, z)
# self._cam_roll = p.readUserDebugParameter(self.roll_slider)
# TODO: recompute view_matrix and proj_matrix only in debug mode
view_matrix1 = p.computeViewMatrixFromYawPitchRoll(
cameraTargetPosition=camera_target_pos,
distance=self._cam_dist,
yaw=self._cam_yaw,
pitch=self._cam_pitch,
roll=self._cam_roll,
upAxisIndex=2)
proj_matrix1 = p.computeProjectionMatrixFOV(
fov=60, aspect=float(RENDER_WIDTH) / RENDER_HEIGHT,
nearVal=0.1, farVal=100.0)
(_, _, px1, _, _) = p.getCameraImage(
width=RENDER_WIDTH, height=RENDER_HEIGHT, viewMatrix=view_matrix1,
projectionMatrix=proj_matrix1, renderer=self.renderer)
rgb_array1 = np.array(px1)
if self.multi_view:
# adding a second camera on the other side of the robot
view_matrix2 = p.computeViewMatrixFromYawPitchRoll(
cameraTargetPosition=(0.316, 0.316, -0.105),
distance=1.05,
yaw=32,
pitch=-13,
roll=0,
upAxisIndex=2)
proj_matrix2 = p.computeProjectionMatrixFOV(
fov=60, aspect=float(RENDER_WIDTH) / RENDER_HEIGHT,
nearVal=0.1, farVal=100.0)
(_, _, px2, _, _) = p.getCameraImage(
width=RENDER_WIDTH, height=RENDER_HEIGHT, viewMatrix=view_matrix2,
projectionMatrix=proj_matrix2, renderer=self.renderer)
rgb_array2 = np.array(px2)
rgb_array_res = np.concatenate((rgb_array1[:, :, :3], rgb_array2[:, :, :3]), axis=2)
else:
rgb_array_res = rgb_array1[:, :, :3]
return rgb_array_res
def _termination(self):
if self.terminated or self._env_step_counter > self.max_steps:
self._observation = self.getExtendedObservation()
return True
return False
def _reward(self):
gripper_pos = self.getArmPos()
distance = np.linalg.norm(self.button_pos - gripper_pos, 2)
# print(distance)
contact_points = p.getContactPoints(self.button_uid, self._kuka.kuka_uid, BUTTON_LINK_IDX)
reward = int(len(contact_points) > 0)
self.n_contacts += reward
contact_with_table = len(p.getContactPoints(self.table_uid, self._kuka.kuka_uid)) > 0
if distance > self._max_distance or contact_with_table:
reward = -1
self.n_steps_outside += 1
else:
self.n_steps_outside = 0
if contact_with_table or self.n_contacts >= N_CONTACTS_BEFORE_TERMINATION \
or self.n_steps_outside >= N_STEPS_OUTSIDE_SAFETY_SPHERE:
self.terminated = True
if self._shape_reward:
if self._is_discrete:
return -distance
else:
# Button pushed
if self.terminated and reward > 0:
return 50
# out of bounds
elif self.terminated and reward < 0:
return -250
# anything else
else:
return -distance
return reward
def __init__(self,
renders=False,
is_discrete=True,
log_folder="robobo_log_folder",
state_dim=-1,
learn_states=False,
srl_model="raw_pixels",
record_data=False,
shape_reward=False,
env_rank=0,
srl_pipe=None,
img_shape=None):
super(RoboboEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self.n_contacts = 0
use_ground_truth = srl_model == 'ground_truth'
use_srl = srl_model != 'raw_pixels'
self.use_srl = use_srl or use_ground_truth
self.use_ground_truth = use_ground_truth
self.use_joints = False
self.relative_pos = RELATIVE_POS
self._is_discrete = is_discrete
self.observation = []
# Start simulation with first observation
self._env_step_counter = 0
self.episode_terminated = False
self.state_dim = state_dim
self._renders = renders
self._shape_reward = shape_reward
self.cuda = th.cuda.is_available()
self.target_pos = None
self.saver = None
if img_shape is None:
self.img_shape = (3, RENDER_HEIGHT, RENDER_WIDTH)
else:
self.img_shape = img_shape
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
action_dim = 2
self._action_bound = 1
action_bounds = np.array([self._action_bound] * action_dim)
self.action_space = spaces.Box(-action_bounds,
action_bounds,
dtype=np.float32)
# SRL model
if self.use_srl:
if use_ground_truth:
self.state_dim = self.getGroundTruthDim()
self.dtype = np.float32
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=self.dtype)
else:
self.dtype = np.uint8
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self.img_shape[2],
self.img_shape[1], 3),
dtype=self.dtype)
if record_data:
print("Recording data...")
self.saver = EpisodeSaver(log_folder,
0,
self.state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states)
# Initialize Baxter effector by connecting to the Gym bridge ROS node:
self.context = zmq.Context()
self.socket = self.context.socket(zmq.PAIR)
self.socket.connect("tcp://{}:{}".format(HOSTNAME, SERVER_PORT))
# note: if takes too long, run first client, then server
print("Waiting for server connection...")
msg = self.socket.recv_json()
print("Connected to server (received message: {})".format(msg))
self.action = [0, 0]
self.reward = 0
self.robobo_pos = np.array([0, 0])
# Initialize the state
if self._renders:
self.image_plot = None
class RoboboEnv(SRLGymEnv):
"""
Robobo robot Environment (Gym wrapper for Robobo Gazebo environment)
The goal of robobo is to go to the location on the table
(signaled with a circle sticker on the table)
:param renders: (bool) Whether to display the GUI or not
:param is_discrete: (bool) true if action space is discrete vs continuous
:param log_folder: (str) name of the folder where recorded data will be stored
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param record_data: (bool) Set to true, record frames with the rewards.
:param shape_reward: (bool) Set to true, reward = -distance_to_goal
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
"""
def __init__(self,
renders=False,
is_discrete=True,
log_folder="robobo_log_folder",
state_dim=-1,
learn_states=False,
srl_model="raw_pixels",
record_data=False,
shape_reward=False,
env_rank=0,
srl_pipe=None,
img_shape=None):
super(RoboboEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self.n_contacts = 0
use_ground_truth = srl_model == 'ground_truth'
use_srl = srl_model != 'raw_pixels'
self.use_srl = use_srl or use_ground_truth
self.use_ground_truth = use_ground_truth
self.use_joints = False
self.relative_pos = RELATIVE_POS
self._is_discrete = is_discrete
self.observation = []
# Start simulation with first observation
self._env_step_counter = 0
self.episode_terminated = False
self.state_dim = state_dim
self._renders = renders
self._shape_reward = shape_reward
self.cuda = th.cuda.is_available()
self.target_pos = None
self.saver = None
if img_shape is None:
self.img_shape = (3, RENDER_HEIGHT, RENDER_WIDTH)
else:
self.img_shape = img_shape
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
action_dim = 2
self._action_bound = 1
action_bounds = np.array([self._action_bound] * action_dim)
self.action_space = spaces.Box(-action_bounds,
action_bounds,
dtype=np.float32)
# SRL model
if self.use_srl:
if use_ground_truth:
self.state_dim = self.getGroundTruthDim()
self.dtype = np.float32
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=self.dtype)
else:
self.dtype = np.uint8
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self.img_shape[2],
self.img_shape[1], 3),
dtype=self.dtype)
if record_data:
print("Recording data...")
self.saver = EpisodeSaver(log_folder,
0,
self.state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states)
# Initialize Baxter effector by connecting to the Gym bridge ROS node:
self.context = zmq.Context()
self.socket = self.context.socket(zmq.PAIR)
self.socket.connect("tcp://{}:{}".format(HOSTNAME, SERVER_PORT))
# note: if takes too long, run first client, then server
print("Waiting for server connection...")
msg = self.socket.recv_json()
print("Connected to server (received message: {})".format(msg))
self.action = [0, 0]
self.reward = 0
self.robobo_pos = np.array([0, 0])
# Initialize the state
if self._renders:
self.image_plot = None
def step(self, action):
"""
:action: (int)
:return: (tensor (np.ndarray)) observation, int reward, bool done, dict extras)
"""
assert self.action_space.contains(action)
# Convert int action to action in (x,y,z) space
self.action = action
self._env_step_counter += 1
# Send the action to the server
self.socket.send_json({"command": "action", "action": self.action})
# Receive state data (position, etc), important to update state related values
self.getEnvState()
# Receive a camera image from the server
self.observation = self.getObservation()
done = self._hasEpisodeTerminated()
if self.saver is not None:
self.saver.step(self.observation, action, self.reward, done,
self.getGroundTruth())
if self.use_srl:
return self.getSRLState(self.observation), self.reward, done, {}
else:
return self.observation, self.reward, done, {}
def getEnvState(self):
"""
Returns a dictionary containing info about the environment state.
:return: (dict) state_data containing data related to the state: target_pos,
robobo_pos and reward.
"""
state_data = self.socket.recv_json()
self.reward = state_data["reward"]
self.target_pos = np.array(state_data["target_pos"])
self.robobo_pos = np.array(state_data["position"])
return state_data
def getObservation(self):
"""
Receive the observation image using a socket
:return: (numpy ndarray) observation
"""
# Receive a camera image from the server
self.observation = recvMatrix(self.socket)
# Resize it:
self.observation = cv2.resize(self.observation,
(self.img_shape[2], self.img_shape[1]),
interpolation=cv2.INTER_AREA)
return self.observation
def getTargetPos(self):
"""
:return (numpy array): Position of the target (button)
"""
return self.target_pos
@staticmethod
def getGroundTruthDim():
"""
:return: (int)
"""
return 2
def getGroundTruth(self):
"""
Alias for getRoboboPos for compatibility between envs
:return: (numpy array)
"""
return np.array(self.getRoboboPos())
def getRoboboPos(self):
"""
:return: ([float])-> np.ndarray Position (x, y, z) of Baxter left gripper
"""
return self.robobo_pos
def reset(self):
"""
Reset the environment
:return: (numpy ndarray) first observation of the env
"""
self.episode_terminated = False
# Step count since episode start
self._env_step_counter = 0
self.socket.send_json({"command": "reset"})
# Update state related variables, important step to get both data and
# metadata that allow reading the observation image
self.getEnvState()
self.observation = self.getObservation()
if self.saver is not None:
self.saver.reset(self.observation, self.getTargetPos(),
self.getGroundTruth())
if self.use_srl:
return self.getSRLState(self.observation)
else:
return self.observation
def _hasEpisodeTerminated(self):
"""
Returns True if the episode is over and False otherwise
"""
if self.episode_terminated or self._env_step_counter > MAX_STEPS:
return True
return False
def closeServerConnection(self):
"""
To be called at the end of running the program, externally
"""
print("Robobo_env client exiting and closing socket...")
self.socket.send_json({"command": "exit"})
self.socket.close()
def render(self, mode='rgb_array'):
"""
:param mode: (str)
:return: (numpy array) BGR image
"""
if mode != "rgb_array":
print('render in human mode not yet supported')
return np.array([])
if self._renders:
plt.ion() # needed for interactive update
if self.image_plot is None:
plt.figure('Robobo RL')
self.image_plot = plt.imshow(bgr2rgb(self.observation),
cmap='gray')
self.image_plot.axes.grid(False)
else:
self.image_plot.set_data(bgr2rgb(self.observation))
plt.draw()
# Wait a bit, so that plot is visible
plt.pause(0.0001)
return self.observation
def __init__(self,
renders=False,
name="Omnirobot",
is_discrete=True,
save_path='srl_zoo/data/',
state_dim=-1,
learn_states=False,
srl_model="raw_pixels",
record_data=False,
action_repeat=1,
random_target=True,
shape_reward=False,
simple_continual_target=False,
circular_continual_move=False,
square_continual_move=False,
eight_continual_move=False,
short_episodes=False,
state_init_override=None,
env_rank=0,
srl_pipe=None,
img_shape=None,
**_):
super(OmniRobotEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
if action_repeat != 1:
raise NotImplementedError
self.server_port = SERVER_PORT + env_rank
self.n_contacts = 0
use_ground_truth = srl_model == 'ground_truth'
use_srl = srl_model != 'raw_pixels'
self.use_srl = use_srl or use_ground_truth
self.use_ground_truth = use_ground_truth
self.use_joints = False
self.relative_pos = RELATIVE_POS
self._is_discrete = is_discrete
self.observation = []
# Start simulation with first observation
self._env_step_counter = 0
self.episode_terminated = False
self.state_dim = state_dim
if img_shape is None:
self.img_shape = (3, RENDER_HEIGHT, RENDER_WIDTH)
else:
self.img_shape = img_shape
self._renders = renders
self._shape_reward = shape_reward
self.cuda = th.cuda.is_available()
self.target_pos = None
self.saver = None
self._random_target = random_target
self.simple_continual_target = simple_continual_target
self.circular_continual_move = circular_continual_move
self.square_continual_move = square_continual_move
self.eight_continual_move = eight_continual_move
self.short_episodes = short_episodes
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
action_dim = 2
self.action_space = RingBox(positive_low=ACTION_POSITIVE_LOW,
positive_high=ACTION_POSITIVE_HIGH,
negative_low=ACTION_NEGATIVE_LOW,
negative_high=ACTION_NEGATIVE_HIGH,
shape=np.array([action_dim]),
dtype=np.float32)
# SRL model
if self.use_srl:
if use_ground_truth:
self.state_dim = self.getGroundTruthDim()
self.dtype = np.float32
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=self.dtype)
else:
self.dtype = np.uint8
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self.img_shape[2],
self.img_shape[1], 3),
dtype=self.dtype)
if record_data:
print("Recording data...")
self.saver = EpisodeSaver(name,
0,
self.state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states,
path=save_path)
if USING_OMNIROBOT_SIMULATOR:
self.socket = OmniRobotSimulatorSocket(
simple_continual_target=simple_continual_target,
circular_continual_move=circular_continual_move,
square_continual_move=square_continual_move,
eight_continual_move=eight_continual_move,
output_size=[self.img_shape[2], self.img_shape[1]],
random_target=self._random_target,
state_init_override=state_init_override)
else:
# Initialize Baxter effector by connecting to the Gym bridge ROS node:
self.context = zmq.Context()
self.socket = self.context.socket(zmq.PAIR)
self.socket.connect("tcp://{}:{}".format(HOSTNAME,
self.server_port))
# note: if takes too long, run first client, then server
print("Waiting for server connection at port {}...".format(
self.server_port))
# hide the output of server
msg = self.socket.recv_json()
print(
"Connected to server on port {} (received message: {})".format(
self.server_port, msg))
self.action = [0, 0]
self.reward = 0
self.robot_pos = np.array([0, 0])
# Initialize the state
if self._renders:
self.image_plot = None
class OmniRobotEnv(SRLGymEnv):
"""
OmniRobot robot Environment (Gym wrapper for OmniRobot environment)
The goal of Omnirobot is to go to the location on the table
(signaled with a circle sticker on the table)
:param renders: (bool) Whether to display the GUI or not
:param is_discrete: (bool) true if action space is discrete vs continuous
:param save_path: (str) name of the folder where recorded data will be stored
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param record_data: (bool) Set to true, record frames with the rewards.
:param shape_reward: (bool) Set to true, reward = -distance_to_goal
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
"""
def __init__(self,
renders=False,
name="Omnirobot",
is_discrete=True,
save_path='srl_zoo/data/',
state_dim=-1,
learn_states=False,
srl_model="raw_pixels",
record_data=False,
action_repeat=1,
random_target=True,
shape_reward=False,
simple_continual_target=False,
circular_continual_move=False,
square_continual_move=False,
eight_continual_move=False,
short_episodes=False,
state_init_override=None,
env_rank=0,
srl_pipe=None,
img_shape=None,
**_):
super(OmniRobotEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
if action_repeat != 1:
raise NotImplementedError
self.server_port = SERVER_PORT + env_rank
self.n_contacts = 0
use_ground_truth = srl_model == 'ground_truth'
use_srl = srl_model != 'raw_pixels'
self.use_srl = use_srl or use_ground_truth
self.use_ground_truth = use_ground_truth
self.use_joints = False
self.relative_pos = RELATIVE_POS
self._is_discrete = is_discrete
self.observation = []
# Start simulation with first observation
self._env_step_counter = 0
self.episode_terminated = False
self.state_dim = state_dim
if img_shape is None:
self.img_shape = (3, RENDER_HEIGHT, RENDER_WIDTH)
else:
self.img_shape = img_shape
self._renders = renders
self._shape_reward = shape_reward
self.cuda = th.cuda.is_available()
self.target_pos = None
self.saver = None
self._random_target = random_target
self.simple_continual_target = simple_continual_target
self.circular_continual_move = circular_continual_move
self.square_continual_move = square_continual_move
self.eight_continual_move = eight_continual_move
self.short_episodes = short_episodes
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
action_dim = 2
self.action_space = RingBox(positive_low=ACTION_POSITIVE_LOW,
positive_high=ACTION_POSITIVE_HIGH,
negative_low=ACTION_NEGATIVE_LOW,
negative_high=ACTION_NEGATIVE_HIGH,
shape=np.array([action_dim]),
dtype=np.float32)
# SRL model
if self.use_srl:
if use_ground_truth:
self.state_dim = self.getGroundTruthDim()
self.dtype = np.float32
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=self.dtype)
else:
self.dtype = np.uint8
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self.img_shape[2],
self.img_shape[1], 3),
dtype=self.dtype)
if record_data:
print("Recording data...")
self.saver = EpisodeSaver(name,
0,
self.state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states,
path=save_path)
if USING_OMNIROBOT_SIMULATOR:
self.socket = OmniRobotSimulatorSocket(
simple_continual_target=simple_continual_target,
circular_continual_move=circular_continual_move,
square_continual_move=square_continual_move,
eight_continual_move=eight_continual_move,
output_size=[self.img_shape[2], self.img_shape[1]],
random_target=self._random_target,
state_init_override=state_init_override)
else:
# Initialize Baxter effector by connecting to the Gym bridge ROS node:
self.context = zmq.Context()
self.socket = self.context.socket(zmq.PAIR)
self.socket.connect("tcp://{}:{}".format(HOSTNAME,
self.server_port))
# note: if takes too long, run first client, then server
print("Waiting for server connection at port {}...".format(
self.server_port))
# hide the output of server
msg = self.socket.recv_json()
print(
"Connected to server on port {} (received message: {})".format(
self.server_port, msg))
self.action = [0, 0]
self.reward = 0
self.robot_pos = np.array([0, 0])
# Initialize the state
if self._renders:
self.image_plot = None
def actionPolicyTowardTarget(self):
"""
:return: (int) action
"""
if abs(self.robot_pos[0] -
self.target_pos[0]) > abs(self.robot_pos[1] -
self.target_pos[1]):
if self._is_discrete:
return int(
Move.FORWARD
) if self.robot_pos[0] < self.target_pos[0] else int(
Move.BACKWARD)
# forward # backward
else:
return DELTA_POS if self.robot_pos[0] < self.target_pos[
0] else -DELTA_POS
else:
if self._is_discrete:
# left # right
return int(
Move.LEFT
) if self.robot_pos[1] < self.target_pos[1] else int(
Move.RIGHT)
else:
return DELTA_POS if self.robot_pos[1] < self.target_pos[
1] else -DELTA_POS
def step(self,
action,
generated_observation=None,
action_proba=None,
action_grid_walker=None):
"""
:param :action: (int)
:param generated_observation:
:param action_proba:
:param action_grid_walker: # Whether or not we want to override the action with the one from a grid walker
:return: (tensor (np.ndarray)) observation, int reward, bool done, dict extras)
"""
if action_grid_walker is None:
action_to_step = action
action_from_teacher = None
else:
action_to_step = action_grid_walker
action_from_teacher = action
assert self.action_space.contains(action_from_teacher)
if not self._is_discrete:
action_to_step = np.array(action_to_step)
assert self.action_space.contains(action_to_step)
# Convert int action to action in (x,y,z) space
# serialize the action
if isinstance(action_to_step, np.ndarray):
self.action = action_to_step.tolist()
elif hasattr(action_to_step,
'dtype'): # convert numpy type to python type
self.action = action.item()
else:
self.action = action_to_step
self._env_step_counter += 1
# Send the action to the server
self.socket.send_json({
"command": "action",
"action": self.action,
"is_discrete": self._is_discrete,
"step_counter": self._env_step_counter
})
# Receive state data (position, etc), important to update state related values
self.getEnvState()
# Receive a camera image from the server
self.observation = self.getObservation(
) if generated_observation is None else generated_observation * 255
done = self._hasEpisodeTerminated()
self.render()
if self.saver is not None:
self.saver.step(self.observation,
action_from_teacher if action_grid_walker
is not None else action_to_step,
self.reward,
done,
self.getGroundTruth(),
action_proba=action_proba)
old_observation = self.getObservation()
if self.use_srl:
return self.getSRLState(
self.observation if generated_observation is None else
old_observation), self.reward, done, {}
else:
return self.observation, self.reward, done, {}
def getEnvState(self):
"""
Returns a dictionary containing info about the environment state.
:return: (dict) state_data containing data related to the state: target_pos,
robot_pos and reward.
"""
state_data = self.socket.recv_json()
self.reward = state_data["reward"]
self.target_pos = np.array(state_data["target_pos"])
self.robot_pos = np.array(state_data["position"])
return state_data
def getObservation(self):
"""
Receive the observation image using a socket
:return: (numpy ndarray) observation
"""
# Receive a camera image from the server
self.observation = recvMatrix(self.socket)
# Resize it:
self.observation = cv2.resize(self.observation,
(self.img_shape[2], self.img_shape[1]),
interpolation=cv2.INTER_AREA)
return self.observation
def getTargetPos(self):
"""
:return (numpy array): Position of the target (button)
"""
return self.target_pos
@staticmethod
def getGroundTruthDim():
"""
:return: (int)
"""
return 2
def getGroundTruth(self):
"""
Alias for getRobotPos for compatibility between envs
:return: (numpy array)
"""
return np.array(self.getRobotPos())
def getRobotPos(self):
"""
:return: ([float])-> np.ndarray Position (x, y, z) of Baxter left gripper
"""
return self.robot_pos
def reset(self,
aligned=False,
init_with_real=False,
generated_observation=None,
state_override=None):
"""
Reset the environment
:param generated_observation:
:param state_override:
:return: (numpy ndarray) first observation of the env
"""
self.episode_terminated = False
# Step count since episode start
self._env_step_counter = 0
# set n contact count
self.n_contacts = 0
self.socket.send_json({
"command": "reset",
"step_counter": self._env_step_counter,
"aligned": aligned
})
# Update state related variables, important step to get both data and
# metadata that allow reading the observation image
self.getEnvState()
if not init_with_real:
self.robot_pos = np.array(
[0, 0]) if state_override is None else state_override
self.observation = self.getObservation(
) if generated_observation is None else generated_observation * 255
if self.saver is not None:
self.saver.reset(self.observation, self.getTargetPos(),
self.getGroundTruth())
old_observation = self.getObservation()
if self.use_srl:
return self.getSRLState(self.observation if generated_observation
is None else old_observation)
else:
return self.observation
def _hasEpisodeTerminated(self):
"""
Returns True if the episode is over and False otherwise
"""
if (self.episode_terminated or self._env_step_counter > MAX_STEPS) or \
(self.n_contacts >= N_CONTACTS_BEFORE_TERMINATION and self.short_episodes and
self.simple_continual_target) or \
(self._env_step_counter > MAX_STEPS_CIRCULAR_TASK_SHORT_EPISODES and self.short_episodes and
self.circular_continual_move):
return True
if np.abs(self.reward -
REWARD_TARGET_REACH) < 0.000001: # reach the target
self.n_contacts += 1
else:
self.n_contacts += 0
return False
def closeServerConnection(self):
"""
To be called at the end of running the program, externally
"""
print("Omnirobot_env client exiting and closing socket...")
self.socket.send_json({"command": "exit"})
self.socket.close()
def render(self, mode='rgb_array'):
"""
:param mode: (str)
:return: (numpy array) BGR image
"""
if self._renders:
if mode != "rgb_array":
print('render in human mode not yet supported')
return np.array([])
plt.ion() # needed for interactive update
if self.image_plot is None:
plt.figure('Omnirobot RL')
self.initVisualizeBoundary()
self.visualizeBoundary()
self.image_plot = plt.imshow(self.observation_with_boundary,
cmap='gray')
self.image_plot.axes.grid(False)
else:
self.visualizeBoundary()
self.image_plot.set_data(self.observation_with_boundary)
plt.draw()
# Wait a bit, so that plot is visible
plt.pause(0.0001)
return self.observation
def initVisualizeBoundary(self):
with open(CAMERA_INFO_PATH, 'r') as stream:
try:
contents = yaml.load(stream)
camera_matrix = np.array(
contents['camera_matrix']['data']).reshape((3, 3))
distortion_coefficients = np.array(
contents['distortion_coefficients']['data']).reshape(
(1, 5))
except yaml.YAMLError as exc:
print(exc)
# camera installation info
r = R.from_euler('xyz', CAMERA_ROT_EULER_COORD_GROUND, degrees=True)
camera_rot_mat_coord_ground = r.as_dcm()
pos_transformer = PosTransformer(camera_matrix,
distortion_coefficients,
CAMERA_POS_COORD_GROUND,
camera_rot_mat_coord_ground)
self.boundary_coner_pixel_pos = np.zeros((2, 4))
# assume that image is undistorted
self.boundary_coner_pixel_pos[:, 0] = \
pos_transformer.phyPosGround2PixelPos([MIN_X, MIN_Y], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos[:, 1] = \
pos_transformer.phyPosGround2PixelPos([MAX_X, MIN_Y], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos[:, 2] = \
pos_transformer.phyPosGround2PixelPos([MAX_X, MAX_Y], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos[:, 3] = \
pos_transformer.phyPosGround2PixelPos([MIN_X, MAX_Y], return_distort_image_pos=False).squeeze()
# transform the corresponding points into cropped image
self.boundary_coner_pixel_pos = self.boundary_coner_pixel_pos - (
np.array(ORIGIN_SIZE) - np.array(CROPPED_SIZE)).reshape(2, 1) / 2.0
# transform the corresponding points into resized image (self.img_shape[2], self.img_shape[1])
self.boundary_coner_pixel_pos[
0, :] *= self.img_shape[2] / CROPPED_SIZE[0]
self.boundary_coner_pixel_pos[
1, :] *= self.img_shape[1] / CROPPED_SIZE[1]
self.boundary_coner_pixel_pos = np.around(
self.boundary_coner_pixel_pos).astype(np.int)
# Create square for vizu of objective in continual square task
if self.square_continual_move:
self.boundary_coner_pixel_pos_continual = np.zeros((2, 4))
# assume that image is undistorted
self.boundary_coner_pixel_pos_continual[:, 0] = \
pos_transformer.phyPosGround2PixelPos([-RADIUS, -RADIUS], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos_continual[:, 1] = \
pos_transformer.phyPosGround2PixelPos([RADIUS, -RADIUS], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos_continual[:, 2] = \
pos_transformer.phyPosGround2PixelPos([RADIUS, RADIUS], return_distort_image_pos=False).squeeze()
self.boundary_coner_pixel_pos_continual[:, 3] = \
pos_transformer.phyPosGround2PixelPos([-RADIUS, RADIUS], return_distort_image_pos=False).squeeze()
# transform the corresponding points into cropped image
self.boundary_coner_pixel_pos_continual = self.boundary_coner_pixel_pos_continual - (
np.array(ORIGIN_SIZE) - np.array(CROPPED_SIZE)).reshape(
2, 1) / 2.0
# transform the corresponding points into resized image (self.img_shape[2], self.img_shape[1])
self.boundary_coner_pixel_pos_continual[
0, :] *= self.img_shape[2] / CROPPED_SIZE[0]
self.boundary_coner_pixel_pos_continual[
1, :] *= self.img_shape[1] / CROPPED_SIZE[1]
self.boundary_coner_pixel_pos_continual = np.around(
self.boundary_coner_pixel_pos_continual).astype(np.int)
elif self.circular_continual_move:
self.center_coordinates = \
pos_transformer.phyPosGround2PixelPos([0, 0], return_distort_image_pos=False).squeeze()
self.center_coordinates = self.center_coordinates - (
np.array(ORIGIN_SIZE) - np.array(CROPPED_SIZE)) / 2.0
# transform the corresponding points into resized image (self.img_shape[2], self.img_shape[1])
self.center_coordinates[0] *= self.img_shape[2] / CROPPED_SIZE[0]
self.center_coordinates[1] *= self.img_shape[1] / CROPPED_SIZE[1]
self.center_coordinates = np.around(
self.center_coordinates).astype(np.int)
# Points to convert radisu in env space
self.boundary_coner_pixel_pos_continual = \
pos_transformer.phyPosGround2PixelPos([0, RADIUS], return_distort_image_pos=False).squeeze()
# transform the corresponding points into cropped image
self.boundary_coner_pixel_pos_continual = self.boundary_coner_pixel_pos_continual - (
np.array(ORIGIN_SIZE) - np.array(CROPPED_SIZE)) / 2.0
# transform the corresponding points into resized image (self.img_shape[2], self.img_shape[1])
self.boundary_coner_pixel_pos_continual[
0] *= self.img_shape[2] / CROPPED_SIZE[0]
self.boundary_coner_pixel_pos_continual[
1] *= self.img_shape[1] / CROPPED_SIZE[1]
self.boundary_coner_pixel_pos_continual = np.around(
self.boundary_coner_pixel_pos_continual).astype(np.int)
def visualizeBoundary(self):
"""
visualize the unvisible boundary, should call initVisualizeBoundary first
"""
self.observation_with_boundary = self.observation.copy()
# Add boundary continual
if self.square_continual_move:
for idx in range(4):
idx_next = idx + 1
cv2.line(
self.observation_with_boundary,
tuple(self.boundary_coner_pixel_pos_continual[:, idx]),
tuple(self.boundary_coner_pixel_pos_continual[:, idx_next %
4]),
(0, 0, 200), 1)
elif self.circular_continual_move:
radius_converted = np.linalg.norm(
self.center_coordinates -
self.boundary_coner_pixel_pos_continual)
cv2.circle(self.observation_with_boundary,
tuple(self.center_coordinates),
np.float32(radius_converted), (0, 0, 200), 2)
# Add boundary of env
for idx in range(4):
idx_next = idx + 1
cv2.line(self.observation_with_boundary,
tuple(self.boundary_coner_pixel_pos[:, idx]),
tuple(self.boundary_coner_pixel_pos[:, idx_next % 4]),
(200, 0, 0), 1)
def __init__(self,
urdf_root=pybullet_data.getDataPath(),
renders=False,
is_discrete=True,
name="mobile_robot",
max_distance=1.6,
shape_reward=False,
record_data=False,
srl_model="raw_pixels",
random_target=False,
force_down=True,
state_dim=-1,
learn_states=False,
verbose=False,
save_path='srl_zoo/data/',
env_rank=0,
srl_pipe=None,
fpv=False,
img_shape=None,
**_):
super(MobileRobotGymEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._timestep = 1. / 240.
self._urdf_root = urdf_root
self._observation = []
self._env_step_counter = 0
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
self._cam_dist = 4.4
self._cam_yaw = 90
self._cam_pitch = -90
self._cam_roll = 0
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._force_down = force_down
self.camera_target_pos = (2, 2, 0)
self._is_discrete = is_discrete
self.terminated = False
self.renderer = p.ER_TINY_RENDERER
self.debug = False
self.n_contacts = 0
self.state_dim = state_dim
self.relative_pos = RELATIVE_POS
self.cuda = th.cuda.is_available()
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(3)
self.robot_uid = None
self.target_pos = np.zeros(3)
self.target_uid = None
# Boundaries of the square env
self._min_x, self._max_x = 0, 4
self._min_y, self._max_y = 0, 4
self.has_bumped = False # Used for handling collisions
self.collision_margin = 0.1
self.walls = None
self.fpv = fpv
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name,
max_distance,
state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states,
path=save_path)
if self._renders:
client_id = p.connect(p.SHARED_MEMORY)
if client_id < 0:
p.connect(p.GUI)
p.resetDebugVisualizerCamera(1.3, 180, -41, [0.52, -0.2, -0.33])
self.debug = True
# Debug sliders for moving the camera
self.x_slider = p.addUserDebugParameter("x_slider", -10, 10,
self.camera_target_pos[0])
self.y_slider = p.addUserDebugParameter("y_slider", -10, 10,
self.camera_target_pos[1])
self.z_slider = p.addUserDebugParameter("z_slider", -10, 10,
self.camera_target_pos[2])
self.dist_slider = p.addUserDebugParameter("cam_dist", 0, 10,
self._cam_dist)
self.yaw_slider = p.addUserDebugParameter("cam_yaw", -180, 180,
self._cam_yaw)
self.pitch_slider = p.addUserDebugParameter(
"cam_pitch", -180, 180, self._cam_pitch)
else:
p.connect(p.DIRECT)
global CONNECTED_TO_SIMULATOR
CONNECTED_TO_SIMULATOR = True
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1,
high=1,
shape=(2, ),
dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
class MobileRobotGymEnv(SRLGymEnv):
"""
Gym wrapper for Mobile Robot environment
WARNING: to be compatible with kuka scripts, additional keyword arguments are discarded
:param urdf_root: (str) Path to pybullet urdf files
:param renders: (bool) Whether to display the GUI or not
:param is_discrete: (bool) Whether to use discrete or continuous actions
:param name: (str) name of the folder where recorded data will be stored
:param max_distance: (float) Max distance between end effector and the button (for negative reward)
:param shape_reward: (bool) Set to true, reward = -distance_to_goal
:param record_data: (bool) Set to true, record frames with the rewards.
:param random_target: (bool) Set the target to a random position
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param verbose: (bool) Whether to print some debug info
:param save_path: (str) location where the saved data should go
:param env_rank: (int) the number ID of the environment
:param pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param fpv: (bool) enable first person view camera
:param srl_model: (str) The SRL_model used
"""
def __init__(self,
urdf_root=pybullet_data.getDataPath(),
renders=False,
is_discrete=True,
name="mobile_robot",
max_distance=1.6,
shape_reward=False,
record_data=False,
srl_model="raw_pixels",
random_target=False,
force_down=True,
state_dim=-1,
learn_states=False,
verbose=False,
save_path='srl_zoo/data/',
env_rank=0,
srl_pipe=None,
fpv=False,
img_shape=None,
**_):
super(MobileRobotGymEnv, self).__init__(srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._timestep = 1. / 240.
self._urdf_root = urdf_root
self._observation = []
self._env_step_counter = 0
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
self._cam_dist = 4.4
self._cam_yaw = 90
self._cam_pitch = -90
self._cam_roll = 0
self._max_distance = max_distance
self._shape_reward = shape_reward
self._random_target = random_target
self._force_down = force_down
self.camera_target_pos = (2, 2, 0)
self._is_discrete = is_discrete
self.terminated = False
self.renderer = p.ER_TINY_RENDERER
self.debug = False
self.n_contacts = 0
self.state_dim = state_dim
self.relative_pos = RELATIVE_POS
self.cuda = th.cuda.is_available()
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(3)
self.robot_uid = None
self.target_pos = np.zeros(3)
self.target_uid = None
# Boundaries of the square env
self._min_x, self._max_x = 0, 4
self._min_y, self._max_y = 0, 4
self.has_bumped = False # Used for handling collisions
self.collision_margin = 0.1
self.walls = None
self.fpv = fpv
self.srl_model = srl_model
if record_data:
self.saver = EpisodeSaver(name,
max_distance,
state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states,
path=save_path)
if self._renders:
client_id = p.connect(p.SHARED_MEMORY)
if client_id < 0:
p.connect(p.GUI)
p.resetDebugVisualizerCamera(1.3, 180, -41, [0.52, -0.2, -0.33])
self.debug = True
# Debug sliders for moving the camera
self.x_slider = p.addUserDebugParameter("x_slider", -10, 10,
self.camera_target_pos[0])
self.y_slider = p.addUserDebugParameter("y_slider", -10, 10,
self.camera_target_pos[1])
self.z_slider = p.addUserDebugParameter("z_slider", -10, 10,
self.camera_target_pos[2])
self.dist_slider = p.addUserDebugParameter("cam_dist", 0, 10,
self._cam_dist)
self.yaw_slider = p.addUserDebugParameter("cam_yaw", -180, 180,
self._cam_yaw)
self.pitch_slider = p.addUserDebugParameter(
"cam_pitch", -180, 180, self._cam_pitch)
else:
p.connect(p.DIRECT)
global CONNECTED_TO_SIMULATOR
CONNECTED_TO_SIMULATOR = True
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1,
high=1,
shape=(2, ),
dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
def getTargetPos(self):
# Return only the [x, y] coordinates
return self.target_pos[:2]
# @staticmethod
def getGroundTruthDim(self):
"""
The new convention for the ground truth (GT): GT should include target position under random target
setting.
"""
## HACK : Monkey-Patch, shit solution to solve problem.
if not self._random_target:
return 2
else:
return 4
def getRobotPos(self):
# Return only the [x, y] coordinates
return np.array(self.robot_pos)[:2]
def getGroundTruth(self):
"""
The new convention for the ground truth (GT): GT should include target position under random target
setting. A better solution would be "change all the environment files, especially srl_env.py" !!!
"""
## HACK: Monkey-Patch, shit solution to solve problem.
robot_pos = self.getRobotPos()
if self._random_target:
if self.relative_pos:
## HACK here! Change srl_env.py and all the other envs in the future !!! TODO TODO TODO
return robot_pos
else:
## check 'envs.observation_space' in rl_baselines/base_classes.py (before self.model.learn) !!!
target_pos = self.getTargetPos()
return np.concatenate([robot_pos, target_pos], axis=0)
else:
if self.relative_pos:
## HACK here! Change srl_env.py and all the other envs in the future !!! TODO TODO TODO
return robot_pos
else:
return robot_pos
def reset(self):
self.terminated = False
p.resetSimulation()
p.setPhysicsEngineParameter(numSolverIterations=150)
p.setTimeStep(self._timestep)
p.loadURDF(os.path.join(self._urdf_root, "plane.urdf"), [0, 0, 0])
p.setGravity(0, 0, -10)
# Init the robot randomly
x_start = self._max_x / 2 + self.np_random.uniform(
-self._max_x / 3, self._max_x / 3)
y_start = self._max_y / 2 + self.np_random.uniform(
-self._max_y / 3, self._max_y / 3)
self.robot_pos = np.array([x_start, y_start, 0])
# Initialize target position
x_pos = 0.9 * self._max_x
y_pos = self._max_y * 3 / 4
if self._random_target:
margin = 0.1 * self._max_x
x_pos = self.np_random.uniform(self._min_x + margin,
self._max_x - margin)
y_pos = self.np_random.uniform(self._min_y + margin,
self._max_y - margin)
self.target_uid = p.loadURDF("/urdf/cylinder.urdf", [x_pos, y_pos, 0],
useFixedBase=True)
self.target_pos = np.array([x_pos, y_pos, 0])
# Add walls
# Path to the urdf file
wall_urdf = "/urdf/wall.urdf"
# Rgba colors
red, green, blue = [0.8, 0, 0, 1], [0, 0.8, 0, 1], [0, 0, 0.8, 1]
wall_left = p.loadURDF(wall_urdf, [self._max_x / 2, 0, 0],
useFixedBase=True)
# Change color
p.changeVisualShape(wall_left, -1, rgbaColor=red)
# getQuaternionFromEuler -> define orientation
wall_bottom = p.loadURDF(wall_urdf, [self._max_x, self._max_y / 2, 0],
p.getQuaternionFromEuler([0, 0, np.pi / 2]),
useFixedBase=True)
wall_right = p.loadURDF(wall_urdf, [self._max_x / 2, self._max_y, 0],
useFixedBase=True)
p.changeVisualShape(wall_right, -1, rgbaColor=green)
wall_top = p.loadURDF(wall_urdf, [self._min_x, self._max_y / 2, 0],
p.getQuaternionFromEuler([0, 0, np.pi / 2]),
useFixedBase=True)
p.changeVisualShape(wall_top, -1, rgbaColor=blue)
self.walls = [wall_left, wall_bottom, wall_right, wall_top]
# Add mobile robot
self.robot_uid = p.loadURDF(os.path.join(self._urdf_root,
"racecar/racecar.urdf"),
self.robot_pos,
useFixedBase=True)
self._env_step_counter = 0
for _ in range(50):
p.stepSimulation()
self._observation = self.getObservation()
if self.saver is not None:
self.saver.reset(self._observation, self.getTargetPos(),
self.getRobotPos())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation)
return np.array(self._observation)
def __del__(self):
if CONNECTED_TO_SIMULATOR:
p.disconnect()
def getObservation(self):
"""
:return: (numpy array)
"""
self._observation = self.render("rgb_array")
return self._observation
def step(self, action):
# True if it has bumped against a wall
self.has_bumped = False
if self._is_discrete:
dv = DELTA_POS
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD)
dx = [-dv, dv, 0, 0][action]
dy = [0, 0, -dv, dv][action]
real_action = np.array([dx, dy])
else:
dv = DELTA_POS
# Add noise to action
dv += self.np_random.normal(0.0, scale=NOISE_STD)
# scale action amplitude between -dv and dv
real_action = np.maximum(np.minimum(action, 1), -1) * dv
if self.verbose:
print(np.array2string(np.array(real_action), precision=2))
previous_pos = self.robot_pos.copy()
self.robot_pos[:2] += real_action
# Handle collisions
for i, (limit, robot_dim) in enumerate(
zip([self._max_x, self._max_y], [ROBOT_LENGTH, ROBOT_WIDTH])):
margin = self.collision_margin + robot_dim / 2
# If it has bumped against a wall, stay at the previous position
if self.robot_pos[i] < margin or self.robot_pos[i] > limit - margin:
self.has_bumped = True
self.robot_pos = previous_pos
break
# Update mobile robot position
p.resetBasePositionAndOrientation(self.robot_uid, self.robot_pos,
[0, 0, 0, 1])
p.stepSimulation()
self._env_step_counter += 1
self._observation = self.getObservation()
reward = self._reward()
done = self._termination()
if self.saver is not None:
self.saver.step(self._observation, action, reward, done,
self.getRobotPos())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation), reward, done, {}
return np.array(self._observation), reward, done, {}
def render(self, mode='human', close=False):
if mode != "rgb_array":
return np.array([])
camera_target_pos = self.camera_target_pos
if self.debug:
self._cam_dist = p.readUserDebugParameter(self.dist_slider)
self._cam_yaw = p.readUserDebugParameter(self.yaw_slider)
self._cam_pitch = p.readUserDebugParameter(self.pitch_slider)
x = p.readUserDebugParameter(self.x_slider)
y = p.readUserDebugParameter(self.y_slider)
z = p.readUserDebugParameter(self.z_slider)
camera_target_pos = (x, y, z)
# TODO: recompute view_matrix and proj_matrix only in debug mode
view_matrix = p.computeViewMatrixFromYawPitchRoll(
cameraTargetPosition=camera_target_pos,
distance=self._cam_dist,
yaw=self._cam_yaw,
pitch=self._cam_pitch,
roll=self._cam_roll,
upAxisIndex=2)
proj_matrix = p.computeProjectionMatrixFOV(fov=60,
aspect=float(self._width) /
self._height,
nearVal=0.1,
farVal=100.0)
(_, _, px1, _, _) = p.getCameraImage(width=self._width,
height=self._height,
viewMatrix=view_matrix,
projectionMatrix=proj_matrix,
renderer=self.renderer)
rgb_array = np.array(px1)
rgb_array_res = rgb_array[:, :, :3]
# if first person view, then stack the obersvation from the car camera
if self.fpv:
# move camera
view_matrix = p.computeViewMatrixFromYawPitchRoll(
cameraTargetPosition=(self.robot_pos[0] - 0.25,
self.robot_pos[1], 0.15),
distance=0.3,
yaw=self._cam_yaw,
pitch=-17,
roll=self._cam_roll,
upAxisIndex=2)
proj_matrix = p.computeProjectionMatrixFOV(
fov=90,
aspect=float(self._width) / self._height,
nearVal=0.1,
farVal=100.0)
# get and stack image
(_, _, px1, _, _) = p.getCameraImage(width=self._width,
height=self._height,
viewMatrix=view_matrix,
projectionMatrix=proj_matrix,
renderer=self.renderer)
rgb_array = np.array(px1)
rgb_array_res = np.dstack([rgb_array_res, rgb_array[:, :, :3]])
return rgb_array_res
def _termination(self):
"""
:return: (bool)
"""
if self.terminated or self._env_step_counter > self.max_steps:
self._observation = self.getObservation()
return True
return False
def _reward(self):
"""
:return: (float)
"""
# Distance to target
distance = np.linalg.norm(self.getTargetPos() - self.robot_pos[:2], 2)
reward = 0
if distance <= REWARD_DIST_THRESHOLD:
reward = 1
# self.terminated = True
# Negative reward when it bumps into a wall
if self.has_bumped:
reward = -1
if self._shape_reward:
return -distance
return reward
class LabyrinthEnv3(SRLGymEnv):
"""
This Labyrinth environment could can at speed 36,000 FPS on 10 CPUs (Intel i9-9900K)
Labyrinth-v3:
add two Palm trees as obstacle (could be destroyed)
add two Oak barrel as obstacle (could be moved and destroyed against wall)
:param name: (str) name of the folder where recorded data will be stored
:param max_distance: (float) Max distance between end effector and the button (for negative reward)
:param record_data: (bool) Set to true, record frames with the rewards.
:param random_target: (bool) Set the target to a random position
:param state_dim: (int) When learning states
:param verbose: (bool) Whether to print some debug info
:param save_path: (str) location where the saved data should go
:param env_rank: (int) the number ID of the environment
:param pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param srl_model: (str) The SRL_model used
"""
def __init__(self,
renders=False,
is_discrete=True,
name="labyrinth",
max_distance=1.6,
shape_reward=False,
record_data=False,
srl_model="raw_pixels",
random_target=False,
state_dim=-1,
verbose=False,
save_path='srl_zoo/data/',
env_rank=0,
srl_pipe=None,
img_shape=None,
**_):
super(LabyrinthEnv3, self).__init__(srl_model=srl_model,
relative_pos=False,
env_rank=env_rank,
srl_pipe=srl_pipe)
self._observation = []
self._renders = renders
self.img_shape = img_shape # channel first !!
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
# self._shape_reward = shape_reward
self._random_target = random_target
self.state_dim = state_dim
self.saver = None
self.verbose = verbose
self.max_steps = MAX_STEPS
self.robot_pos = np.zeros(2, dtype=int) # index on the self.map array
self.target_pos = [
] # list of indexes (target positions) on the self.map array
self.previous_robot_pos = None
self.has_bumped = False # Used for handling collisions
self.count_collected_tresors = 0
self._env_step_counter = 0
self.srl_model = srl_model
self.maze_size = MAZE_SIZE
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
if record_data:
self.saver = EpisodeSaver(name,
max_distance,
state_dim,
globals_=getGlobals(),
path=save_path)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
def create_map(self):
"""
-1 is wall, 0 is free space, 1 is key(tresor), 2 is robot, 3 is obstacle (Palm tree)
4 is oak
"""
# Create map without robot
# put walls
self.map = np.zeros((self.maze_size, self.maze_size))
self.map[:, 0] = -1
self.map[:, -1] = -1
self.map[0, :] = -1
self.map[-1, :] = -1
self.map[:3, 4] = -1
self.map[-3:, 4] = -1
# put tresors (targets)
self.target_pos = []
self.obstacle_pos = []
self.oak_pos = []
if not self._random_target:
self.target_pos.append(np.array([1, -2], dtype=int))
self.target_pos.append(np.array([-2, 1], dtype=int))
self.obstacle_pos.append(np.array([2, 2], dtype=int))
else:
valid_target_pos_list = []
for i in range(self.maze_size):
for j in range(self.maze_size):
if self.map[i, j] == 0:
valid_target_pos_list.append(
np.array([i, j], dtype=int))
obstacle_ind = self.np_random.choice(np.arange(
len(valid_target_pos_list)),
2,
replace=False)
for index in sorted(obstacle_ind, reverse=True):
self.obstacle_pos.append(valid_target_pos_list.pop(index))
oak_ind = self.np_random.choice(np.arange(
len(valid_target_pos_list)),
2,
replace=False)
for index in sorted(oak_ind, reverse=True):
self.oak_pos.append(valid_target_pos_list.pop(index))
for index in self.np_random.choice(
np.arange(len(valid_target_pos_list)), 2,
replace=False): # randomly choose two targets
self.target_pos.append(valid_target_pos_list[index])
for key_pos in self.target_pos:
self.map[key_pos[0], key_pos[1]] = 1
for key_pos in self.obstacle_pos:
self.map[key_pos[0], key_pos[1]] = 3
for key_pos in self.oak_pos:
self.map[key_pos[0], key_pos[1]] = 4
valid_robot_pos_list = []
for i in range(self.maze_size):
for j in range(self.maze_size):
if self.map[i, j] == 0:
valid_robot_pos_list.append(np.array([i, j], dtype=int))
# load images
target_img_ori = cv2.imread("./environments/labyrinth/tresors_128.png")
robot_img_ori = cv2.imread("./environments/labyrinth/corsair_128.png")
palm_img_ori = cv2.imread("./environments/labyrinth/palm_128.jpg")
oak_img_ori = cv2.imread("./environments/labyrinth/oak_128.jpg")
map_h, map_w = self.map.shape
self.square_size = int(self._height / map_h)
self.target_img = cv2.resize(
target_img_ori, (self.square_size, self.square_size))[..., ::-1]
self.robot_img = cv2.resize(
robot_img_ori, (self.square_size, self.square_size))[..., ::-1]
self.palm_img = cv2.resize(
palm_img_ori, (self.square_size, self.square_size))[..., ::-1]
self.oak_img = cv2.resize(
oak_img_ori, (self.square_size, self.square_size))[..., ::-1]
return valid_robot_pos_list
def getGroundTruthDim(self):
"""
The new convention for the ground truth (GT): GT should include target position under random target
setting.
"""
return self.maze_size**2
def getGroundTruth(self):
return self.map.ravel()
def reset(self):
self.count_collected_tresors = 0
self._env_step_counter = 0
self.has_bumped = False
self.previous_robot_pos = None
self._observation = None
valid_robot_pos_list = self.create_map()
# put robot
self.robot_pos = valid_robot_pos_list[self.np_random.choice(
np.arange(len(valid_robot_pos_list)))]
self.map[self.robot_pos[0], self.robot_pos[1]] = 2
self._observation = self.getObservation(start=True)
if self.saver is not None:
self.saver.reset(self._observation, np.zeros(2), np.zeros(2))
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation)
return self._observation
def getObservation(self, start=False):
"""
:return: (numpy array)
"""
self._observation = self.render("rgb_array", start=start)
return self._observation
def valid_action(self, action):
"""
The most important function. It's define the rule of Labyrinth
check whether an action is valid
return (bool)
"""
if action == 0:
real_action = np.array([1, 0])
elif action == 1:
real_action = np.array([0, 1])
elif action == 2:
real_action = np.array([-1, 0])
elif action == 3:
real_action = np.array([0, -1])
else:
raise NotImplementedError
next_pos = self.robot_pos + real_action
# import ipdb; ipdb.set_trace()
valid = np.all(next_pos >= 0) and np.all(
next_pos < self.maze_size) and (self.map[next_pos[0], next_pos[1]]
!= -1)
return valid, next_pos, real_action
def step(self, action):
## action = 0, 1, 2, 3
# True if it has bumped against a wall
self._env_step_counter += 1
self.has_bumped = False
previous_pos = self.robot_pos.copy()
self.previous_robot_pos = previous_pos
valid, next_pos, real_action = self.valid_action(action)
self.has_bumped = not valid
if self.has_bumped:
reward = -1
elif (self.map[next_pos[0], next_pos[1]] == 1):
self.count_collected_tresors += 1
reward = 10
elif (self.map[next_pos[0], next_pos[1]] == 4):
next_next_pos = next_pos + real_action
if self.map[next_next_pos[0], next_next_pos[1]] == -1: # it's wall
pass # nothing change
else:
self.map[next_next_pos[0], next_next_pos[1]] = 4
reward = -0.1
else:
reward = -0.1
done = (self.count_collected_tresors == len(
self.target_pos)) or (self._env_step_counter > self.max_steps)
if self.has_bumped:
# no need to update observation
pass
else:
self.map[previous_pos[0], previous_pos[1]] = 0 # free space
# update robot position and the map
self.robot_pos = next_pos
self.map[self.robot_pos[0], self.robot_pos[1]] = 2
# update observation
self._observation = self.getObservation()
if self.saver is not None:
# HACK TODO TODO: used for the estimation of GTC (ground truth correlation)
if reward >= 1:
discret_reward = 1
elif reward <= -1:
discret_reward = -1
else:
discret_reward = 0
self.saver.step(self._observation, action, discret_reward, done,
np.zeros(2))
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation), reward, done, {}
return self._observation, reward, done, {}
def render(self, mode='rgb_array', start=False):
# [OPT] this part could be optimized again if needed, but it's not necessary, because 36,000 FPS
# has already been ~5 times faster then 7000 FPS of PPO2 (with 10 CPUs).
if self._observation is None or (isinstance(self._observation, list)
and len(
self._observation) == 0) or start:
previous_obs = 255 * np.ones(
(self._height, self._width, 3), dtype=np.uint8)
else:
previous_obs = self._observation
map_h, map_w = self.map.shape
square_size = self.square_size
for i in range(map_h):
for j in range(map_w):
if self.map[i, j] == -1:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) * square_size, :] = 0
elif self.map[i, j] == 1:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) *
square_size, :] = self.target_img
elif self.map[i, j] == 2:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) *
square_size, :] = self.robot_img
elif self.map[i, j] == 3:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) *
square_size, :] = self.palm_img
elif self.map[i, j] == 4:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) *
square_size, :] = self.oak_img
elif self.map[i, j] == 0:
# print(self.previous_robot_pos)
if self.previous_robot_pos is not None and i == self.previous_robot_pos[
0] and j == self.previous_robot_pos[1]:
previous_obs[i * square_size:(i + 1) * square_size,
j * square_size:(j + 1) *
square_size, :] = 255
else:
pass
else:
raise NotImplementedError
if self._renders:
cv2.imshow(
"Keep pressing (any key) to display or 'q'/'Esc' to quit.",
previous_obs[..., ::-1])
key = cv2.waitKey(0)
if key == ord('q') or key == 27: # 'q' or 'Esc'
cv2.destroyAllWindows()
raise KeyboardInterrupt
return previous_obs
def interactive(self, show_map=False):
image = self.getObservation(start=True)
cv2.imshow("image", image[..., ::-1])
while True:
k = cv2.waitKey(0)
if k == 27 or k == ord("q"): # press 'Esc' or 'q' to quit
break
elif k == ord("2") or k == 84:
# down
action = 0
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("6") or k == 83:
# right
action = 1
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("8") or k == 82:
# up
action = 2
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
elif k == ord("4") or k == 81:
# left
action = 3
image, reward, done, _ = self.step(action)
print("Action: {}, Reward: {}, Done: {}".format(
action, reward, done))
if show_map:
print(self.map)
cv2.imshow("image", image[..., ::-1])
else:
print("You are pressing the key: {}".format(k))
class CarRacingEnv(GymCarRacing, SRLGymEnv):
def __init__(self,
name="car_racing",
renders=False,
record_data=False,
is_discrete=True,
state_dim=-1,
learn_states=False,
save_path='srl_zoo/data/',
srl_model="raw_pixels",
env_rank=0,
srl_pipe=None,
lookahead=5,
**_):
"""
Gym wrapper for Racing car environment
WARNING: to be compatible with kuka scripts, additional keyword arguments are discarded
:param name: (str) name of the folder where recorded data will be stored
:param renders: (bool) Whether to display the GUI or not
:param record_data: (bool) Set to true, record frames with the rewards.
:param is_discrete: (bool) Whether to use discrete or continuous actions
:param state_dim: (int) When learning states
:param learn_states: (bool)
:param save_path: (str) location where the saved data should go
:param srl_model: (str) The SRL_model used
:param env_rank: (int) the number ID of the environment
:param srl_pipe: (Queue, [Queue]) contains the input and output of the SRL model
:param lookahead: (int) How many segments ahead of the current position of the track should the target be
"""
SRLGymEnv.__init__(self,
srl_model=srl_model,
relative_pos=RELATIVE_POS,
env_rank=env_rank,
srl_pipe=srl_pipe,
img_shape=None)
GymCarRacing.__init__(self)
self._renders = renders
self.img_shape = img_shape
if self.img_shape is None:
self._width = RENDER_WIDTH
self._height = RENDER_HEIGHT
else:
self._height, self._width = self.img_shape[1:]
self._is_discrete = is_discrete
self.lookahead = lookahead
self.relative_pos = RELATIVE_POS
self._env_step_counter = 0
self._observation = None
self.saver = None
if record_data:
self.saver = EpisodeSaver(name,
None,
state_dim,
globals_=getGlobals(),
relative_pos=RELATIVE_POS,
learn_states=learn_states,
path=save_path)
# Accelerate, brake, stear left, stear right
if self._is_discrete:
self.action_space = spaces.Discrete(N_DISCRETE_ACTIONS)
else:
self.action_space = spaces.Box(low=-1,
high=1,
shape=(3, ),
dtype=np.float32)
if self.srl_model == "ground_truth":
self.state_dim = self.getGroundTruthDim()
if self.srl_model == "raw_pixels":
self.observation_space = spaces.Box(low=0,
high=255,
shape=(self._height,
self._width, 3),
dtype=np.uint8)
else:
self.observation_space = spaces.Box(low=-np.inf,
high=np.inf,
shape=(self.state_dim, ),
dtype=np.float32)
def getTargetPos(self):
# get the nearest track segment to the current position
# then return the track segment position that is ahead of the nearest track segement
nearest_idx = np.argmin(
list(
map(
lambda a: np.sqrt(
np.sum(
((list(a[2:4]) + [0, 0, 0]) - self.getGroundTruth(
))**2)), self.track)))
return np.array(
list(self.track[(nearest_idx + self.lookahead) %
len(self.track)][2:4]) + [0, 0, 0])
@staticmethod
def getGroundTruthDim():
return 5
def getGroundTruth(self):
# the car's current x,y position, angle, speed and angular speed
return np.array(
list(self.car.__dict__["hull"].position) + [
self.car.__dict__["hull"].angle, self.car.__dict__["hull"].
inertia, self.car.__dict__["hull"].angularVelocity
])
def getObservation(self):
"""
:return: (numpy array)
"""
self._observation = self.render("rgb_array")
self._observation = cv2.resize(self._observation,
(self._width, self._height))
return self._observation
def reset(self):
super().reset()
self._env_step_counter = 0
self.getObservation()
if self.saver is not None:
self.saver.reset(self._observation, self.getTargetPos(),
self.getGroundTruth())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation)
return self._observation
def step(self, action):
if action is not None:
if self._is_discrete:
self.car.steer([-1, 1, 0, 0][action])
self.car.gas([0, 0, 1, 0][action])
self.car.brake([0, 0, 0, 1][action])
else:
self.car.steer(-action[0])
self.car.gas(action[1])
self.car.brake(action[2])
self.car.step(1.0 / FPS)
self.world.Step(1.0 / FPS, 6 * 30, 2 * 30)
self.t += 1.0 / FPS
self.getObservation()
step_reward = 0
done = False
if action is not None: # First step without action, called from reset()
self.reward -= 0.1
self._env_step_counter += 1
# We actually don't want to count fuel spent, we want car to be faster.
self.car.fuel_spent = 0.0
step_reward = self.reward - self.prev_reward
self.prev_reward = self.reward
if self.tile_visited_count == len(
self.track) or self._env_step_counter >= MAX_STEPS:
done = True
x, y = self.car.hull.position
if abs(x) > PLAYFIELD or abs(y) > PLAYFIELD:
done = True
step_reward = -100
if self.saver is not None:
self.saver.step(self._observation, action, step_reward, done,
self.getGroundTruth())
if self.srl_model != "raw_pixels":
return self.getSRLState(self._observation), step_reward, done, {}
return np.array(self._observation), step_reward, done, {}
# Copied for the original Gym Racing Car env, it is modified to be able to remove the render window.
def render(self, mode='human'):
if self.viewer is None:
self.viewer = rendering.Viewer(WINDOW_W, WINDOW_H)
self.viewer.window.set_visible(self._renders)
self.score_label = pyglet.text.Label('0000',
font_size=36,
x=20,
y=WINDOW_H * 2.5 / 40.00,
anchor_x='left',
anchor_y='center',
color=(255, 255, 255, 255))
self.transform = rendering.Transform()
if "t" not in self.__dict__:
return # reset() not called yet
zoom = 0.1 * SCALE * max(1 - self.t, 0) + ZOOM * SCALE * min(
self.t, 1) # Animate zoom first second
scroll_x = self.car.hull.position[0]
scroll_y = self.car.hull.position[1]
angle = -self.car.hull.angle
vel = self.car.hull.linearVelocity
if np.linalg.norm(vel) > 0.5:
angle = math.atan2(vel[0], vel[1])
self.transform.set_scale(zoom, zoom)
self.transform.set_translation(
WINDOW_W / 2 - (scroll_x * zoom * math.cos(angle) -
scroll_y * zoom * math.sin(angle)),
WINDOW_H / 4 - (scroll_x * zoom * math.sin(angle) +
scroll_y * zoom * math.cos(angle)))
self.transform.set_rotation(angle)
self.car.draw(self.viewer, mode != "state_pixels")
arr = None
win = self.viewer.window
if mode != 'state_pixels':
win.switch_to()
win.dispatch_events()
if mode == "rgb_array" or mode == "state_pixels":
win.clear()
t = self.transform
if mode == 'rgb_array':
VP_W = VIDEO_W
VP_H = VIDEO_H
else:
VP_W = STATE_W
VP_H = STATE_H
gl.glViewport(0, 0, VP_W, VP_H)
t.enable()
self.render_road()
for geom in self.viewer.onetime_geoms:
geom.render()
t.disable()
self.render_indicators(WINDOW_W,
WINDOW_H) # TODO: find why 2x needed, wtf
image_data = pyglet.image.get_buffer_manager().get_color_buffer(
).get_image_data()
arr = np.fromstring(image_data.data, dtype=np.uint8, sep='')
arr = arr.reshape(VP_H, VP_W, 4)
arr = arr[::-1, :, 0:3]
# agent can call or not call env.render() itself when recording video.
if mode == "rgb_array" and not self.human_render:
win.flip()
if mode == 'human':
self.human_render = True
win.clear()
t = self.transform
gl.glViewport(0, 0, WINDOW_W, WINDOW_H)
t.enable()
self.render_road()
for geom in self.viewer.onetime_geoms:
geom.render()
t.disable()
self.render_indicators(WINDOW_W, WINDOW_H)
win.flip()
self.viewer.onetime_geoms = []
return arr
class RLBenchDSEnv():
"""
wrapped_env is the rlbench env, passed in initialized. This class just calls actions on that env.
"""
def __init__(self,
name="RLBenchDS-unset",
wrapped_env=None,
max_steps_per_epoch=100,
path=None,
**_kwargs):
#Passed in in dataset_generator.py kwargs.
self.env = wrapped_env
self.max_steps_per_epoch = max_steps_per_epoch
print("Recording data...")
try:
env_name = self.env.task.get_name()
except:
env_name = self.env.unwrapped.spec.id
print(env_name)
self.saver = EpisodeSaver(name, env_name=env_name, path=path)
self.action_space = self.env.action_space
def seed(self, seed=None):
"""
Seed random generator
:param seed: (int)
:return: ([int])
"""
self.np_random, seed = seeding.np_random(seed)
return [seed]
def step(self, action, t=0):
"""
:action: (int)
:return: (observation, int reward, bool done, dict extras)
"""
self.action = action
self.observation, reward, done, _ = self.env.step(action)
ground_truth, obs = self.observation['state'], self.observation[
'observation']
done = done or t == self.max_steps_per_epoch
self.saver.step(obs, action, reward, done, ground_truth)
return self.observation, reward, done, {}
@staticmethod
def getGroundTruthDim():
"""
:return: (int)
"""
return None
def getGroundTruth(self):
"""
Don't need this method
:return: (numpy array)
"""
return None
def reset(self):
"""
Reset the environment
:return: (numpy ndarray) first observation of the env
"""
self.observation = self.env.reset()
ground_truth, obs = self.observation['state'], self.observation[
'observation']
if self.saver is not None:
self.saver.reset(obs, ground_truth)
return self.observation
def render(self, mode='rgb_array'):
"""
:param mode: (str)
:return: (numpy array) BGR image
"""
return None
