class Atlas(gym.Env):
"""
Y axis is the vertical axis.
Base class for Webots actors in a Scene.
These environments create single-player scenes and behave like normal Gym environments, if
you don't use multiplayer.
"""
electricity_cost = -2.0 # cost for using motors -- this parameter should be carefully tuned against reward for making progress, other values less improtant
stall_torque_cost = -0.1 # cost for running electric current through a motor even at zero rotational speed, small
foot_collision_cost = -1.0 # touches another leg, or other objects, that cost makes robot avoid smashing feet into itself
joints_at_limit_cost = -0.2 # discourage stuck joints
episode_reward = 0
frame = 0
_max_episode_steps = 1000
initial_y = None
body_xyz = None
joint_angles = None
joint_exceed_limit = False
ignore_frame = 1
def __init__(self, action_dim, obs_dim):
self.robot = Supervisor()
solid_def_names = self.read_all_def()
self.def_node_field_list = self.get_all_fields(solid_def_names)
self.robot_node = self.robot.getFromDef('Atlas')
self.boom_base = self.robot.getFromDef('BoomBase')
self.boom_base_trans_field = self.boom_base.getField("translation")
self.timeStep = int(self.robot.getBasicTimeStep() * self.ignore_frame) # ms
self.find_and_enable_devices()
high = np.ones([action_dim])
self.action_space = gym.spaces.Box(-high, high, dtype=np.float32)
high = np.inf*np.ones([obs_dim])
self.observation_space = gym.spaces.Box(-high, high, dtype=np.float32)
def read_all_def(self, file_name ='../../worlds/atlas_change_foot.wbt'):
no_solid_str_list = ['HingeJoint', 'BallJoint', 'Hinge2Joint', 'Shape', 'Group', 'Physics']
with open(file_name) as f:
content = f.readlines()
# you may also want to remove whitespace characters like `\n` at the end of each line
content = [x.strip() for x in content]
def_str_list = []
for x in content:
if 'DEF' in x:
def_str_list.append(x)
for sub_str in no_solid_str_list:
for i in range(len(def_str_list)):
if sub_str in def_str_list[i]:
def_str_list[i] = 'remove_def'
solid_def_names = []
for def_str in def_str_list:
if 'remove_def' != def_str:
def_str_temp_list = def_str.split()
solid_def_names.append(def_str_temp_list[def_str_temp_list.index('DEF') + 1])
print(solid_def_names)
print('There are duplicates: ',len(solid_def_names) != len(set(solid_def_names)))
return solid_def_names
def get_all_fields(self, solid_def_names):
def_node_field_list = []
for def_name in solid_def_names:
def_node = self.robot.getFromDef(def_name)
node_trans_field = def_node.getField("translation")
node_rot_field = def_node.getField("rotation")
# print(def_name)
node_ini_trans = node_trans_field.getSFVec3f()
node_ini_rot = node_rot_field.getSFRotation()
def_node_field_list.append({'def_node': def_node,
'node_trans_field': node_trans_field,
'node_rot_field': node_rot_field,
'node_ini_trans': node_ini_trans,
'node_ini_rot': node_ini_rot})
return def_node_field_list
def reset_all_fields(self):
for def_node_field in self.def_node_field_list:
def_node_field['node_trans_field'].setSFVec3f(def_node_field['node_ini_trans'])
def_node_field['node_rot_field'].setSFRotation(def_node_field['node_ini_rot'])
def find_and_enable_devices(self):
# inertial unit
self.inertial_unit = self.robot.getInertialUnit("inertial unit")
self.inertial_unit.enable(self.timeStep)
# gps
self.gps = self.robot.getGPS("gps")
self.gps.enable(self.timeStep)
# foot sensors
self.fsr = [self.robot.getTouchSensor("RFsr"), self.robot.getTouchSensor("LFsr")]
for i in range(len(self.fsr)):
self.fsr[i].enable(self.timeStep)
# all motors
# motor_names = [# 'HeadPitch', 'HeadYaw',
# 'LLegUay', 'LLegLax', 'LLegKny',
# 'LLegLhy', 'LLegMhx', 'LLegUhz',
# 'RLegUay', 'RLegLax', 'RLegKny',
# 'RLegLhy', 'RLegMhx', 'RLegUhz',
# ]
motor_names = [ # 'HeadPitch', 'HeadYaw',
'BackLbz', 'BackMby', 'BackUbx', 'NeckAy',
'LLegLax', 'LLegMhx', 'LLegUhz',
'RLegLax', 'RLegMhx', 'RLegUhz',
]
self.motors = []
for i in range(len(motor_names)):
self.motors.append(self.robot.getMotor(motor_names[i]))
# leg pitch motors
self.legPitchMotor = [self.robot.getMotor('RLegLhy'),
self.robot.getMotor('RLegKny'),
self.robot.getMotor('RLegUay'),
self.robot.getMotor('LLegLhy'),
self.robot.getMotor('LLegKny'),
self.robot.getMotor('LLegUay')]
for i in range(len(self.legPitchMotor)):
self.legPitchMotor[i].enableTorqueFeedback(self.timeStep)
# leg pitch sensors
self.legPitchSensor =[self.robot.getPositionSensor('RLegLhyS'),
self.robot.getPositionSensor('RLegKnyS'),
self.robot.getPositionSensor('RLegUayS'),
self.robot.getPositionSensor('LLegLhyS'),
self.robot.getPositionSensor('LLegKnyS'),
self.robot.getPositionSensor('LLegUayS')]
for i in range(len(self.legPitchSensor)):
self.legPitchSensor[i].enable(self.timeStep)
def apply_action(self, a):
assert (np.isfinite(a).all())
for n, j in enumerate(self.legPitchMotor):
max_joint_angle = j.getMaxPosition()
min_joint_angle = j.getMinPosition()
mean_angle = 0.5 * (max_joint_angle + min_joint_angle)
half_range_angle = 0.5 * (max_joint_angle - min_joint_angle)
j.setPosition(mean_angle + half_range_angle * float(np.clip(a[n], -1, +1)))
# joint_angle = self.read_joint_angle(joint_idx=n)
# torque = 0.5 * j.getMaxTorque() * float(np.clip(a[n], -1, +1))
# if joint_angle > max_joint_angle:
# j.setPosition(max_joint_angle - 0.1)
# # j.setTorque(-1.0 * abs(torque))
# elif joint_angle < min_joint_angle:
# j.setPosition(min_joint_angle + 0.1)
# # j.setTorque(abs(torque))
# else:
# j.setTorque(torque)
def read_joint_angle(self, joint_idx):
joint_angle = self.legPitchSensor[joint_idx].getValue() % (2.0 * np.pi)
if joint_angle > np.pi:
joint_angle -= 2.0 * np.pi
max_joint_angle = self.legPitchMotor[joint_idx].getMaxPosition()
min_joint_angle = self.legPitchMotor[joint_idx].getMinPosition()
if joint_angle > max_joint_angle + 0.05 or joint_angle < min_joint_angle - 0.05:
self.joint_exceed_limit = True
return joint_angle
def calc_state(self):
joint_states = np.zeros(2*len(self.legPitchMotor))
# even elements [0::2] position, scaled to -1..+1 between limits
for r in range(6):
joint_angle = self.read_joint_angle(joint_idx=r)
if r in [0, 3]:
joint_states[2 * r] = (-joint_angle - np.deg2rad(35)) / np.deg2rad(80)
elif r in [1, 4]:
joint_states[2 * r] = 1 - joint_angle / np.deg2rad(75)
elif r in [2, 5]:
joint_states[2 * r] = -joint_angle / np.deg2rad(45)
# odd elements [1::2] angular speed, scaled to show -1..+1
for r in range(6):
if self.joint_angles is None:
joint_states[2 * r + 1] = 0.0
else:
joint_states[2 * r + 1] = 0.5 * (joint_states[2*r] - self.joint_angles[r])
self.joint_angles = np.copy(joint_states[0::2])
self.joint_speeds = joint_states[1::2]
self.joints_at_limit = np.count_nonzero(np.abs(joint_states[0::2]) > 0.99)
self.joint_torques = np.zeros(len(self.legPitchMotor))
for i in range(len(self.legPitchMotor)):
self.joint_torques[i] = self.legPitchMotor[i].getTorqueFeedback() \
/ self.legPitchMotor[i].getAvailableTorque()
if self.body_xyz is None:
self.body_xyz = np.asarray(self.gps.getValues())
self.body_speed = np.zeros(3)
else:
self.body_speed = (np.asarray(self.gps.getValues()) - self.body_xyz) / (self.timeStep * 1e-3)
self.body_xyz = np.asarray(self.gps.getValues())
body_local_speed = np.copy(self.body_speed)
body_local_speed[0], body_local_speed[2] = self.calc_local_speed()
# print('speed: ', np.linalg.norm(self.body_speed))
y = self.body_xyz[1]
if self.initial_y is None:
self.initial_y = y
self.body_rpy = self.inertial_unit.getRollPitchYaw()
'''
The roll angle indicates the unit's rotation angle about its x-axis,
in the interval [-π,π]. The roll angle is zero when the InertialUnit is horizontal,
i.e., when its y-axis has the opposite direction of the gravity (WorldInfo defines
the gravity vector).
The pitch angle indicates the unit's rotation angle about is z-axis,
in the interval [-π/2,π/2]. The pitch angle is zero when the InertialUnit is horizontal,
i.e., when its y-axis has the opposite direction of the gravity.
If the InertialUnit is placed on the Robot with a standard orientation,
then the pitch angle is negative when the Robot is going down,
and positive when the robot is going up.
The yaw angle indicates the unit orientation, in the interval [-π,π],
with respect to WorldInfo.northDirection.
The yaw angle is zero when the InertialUnit's x-axis is aligned with the north direction,
it is π/2 when the unit is heading east, and -π/2 when the unit is oriented towards the west.
The yaw angle can be used as a compass.
'''
more = np.array([
y - self.initial_y,
0, 0,
0.3 * body_local_speed[0], 0.3 * body_local_speed[1], 0.3 * body_local_speed[2],
# 0.3 is just scaling typical speed into -1..+1, no physical sense here
self.body_rpy[0] / np.pi, self.body_rpy[1] / np.pi], dtype=np.float32)
self.feet_contact = np.zeros(2)
for j in range(len(self.fsr)):
self.feet_contact[j] = self.fsr[j].getValue()
return np.clip(np.concatenate([more] + [joint_states] + [self.feet_contact]), -5, +5)
def calc_local_speed(self):
'''
# progress in potential field is speed*dt, typical speed is about 2-3 meter per second,
this potential will change 2-3 per frame (not per second),
# all rewards have rew/frame units and close to 1.0
'''
direction_r = self.body_xyz - np.asarray(self.boom_base_trans_field.getSFVec3f())
# print('robot_xyz: {}, boom_base_xyz: {}'.format(self.body_xyz,
# np.asarray(self.boom_base_trans_field.getSFVec3f())))
direction_r = direction_r[[0, 2]] / np.linalg.norm(direction_r[[0, 2]])
direction_t = np.dot(np.asarray([[0, 1],
[-1, 0]]), direction_r.reshape((-1, 1)))
return np.dot(self.body_speed[[0, 2]], direction_t), np.dot(self.body_speed[[0, 2]], direction_r)
def alive_bonus(self, y, pitch):
return +1 if abs(y) > 0.5 and abs(pitch) < 1.0 else -1
def render(self, mode='human'):
file_name = 'img.jpg'
self.robot.exportImage(file=file_name, quality=100)
return cv2.imread(file_name, -1)
def step(self, action):
for i in range(self.ignore_frame):
self.apply_action(action)
simulation_state = self.robot.step(self.timeStep)
state = self.calc_state() # also calculates self.joints_at_limit
# state[0] is body height above ground, body_rpy[1] is pitch
alive = float(self.alive_bonus(state[0] + self.initial_y,
self.body_rpy[1]))
progress, _ = self.calc_local_speed()
# print('progress: {}'.format(progress))
feet_collision_cost = 0.0
'''
let's assume we have DC motor with controller, and reverse current braking
'''
electricity_cost = self.electricity_cost * float(np.abs(
self.joint_torques * self.joint_speeds).mean())
electricity_cost += self.stall_torque_cost * float(np.square(self.joint_torques).mean())
joints_at_limit_cost = float(self.joints_at_limit_cost * self.joints_at_limit)
rewards = [
alive,
progress,
electricity_cost,
joints_at_limit_cost,
feet_collision_cost
]
self.episode_reward += progress
self.frame += 1
done = (-1 == simulation_state) or (self._max_episode_steps <= self.frame) \
or (alive < 0) or (not np.isfinite(state).all())
# print('frame: {}, alive: {}, done: {}, body_xyz: {}'.format(self.frame, alive, done, self.body_xyz))
# print('state_{} \n action_{}, reward_{}'.format(state, action, sum(rewards)))
return state, sum(rewards), done, {}
def run(self):
# Main loop.
for i in range(self._max_episode_steps):
action = np.random.uniform(-1, 1, 6)
state, reward, done, _ = self.step(action)
# print('state_{} \n action_{}, reward_{}'.format(state, action, reward))
if done:
break
def reset(self, is_eval_only = False):
self.initial_y = None
self.body_xyz = None
self.joint_angles = None
self.frame = 0
self.episode_reward = 0
self.joint_exceed_limit = False
for i in range(100):
for j in self.motors:
j.setPosition(0)
for k in range(len(self.legPitchMotor)):
j = self.legPitchMotor[k]
j.setPosition(0)
self.robot.step(self.timeStep)
self.robot.simulationResetPhysics()
self.reset_all_fields()
for i in range(10):
self.robot_node.moveViewpoint()
self.robot.step(self.timeStep)
if is_eval_only:
time.sleep(1)
return self.calc_state()
class WebotsKukaEnv(gym.Env):
metadata = {'render.modes': ['human']}
def __init__(self):
self.alpha = alpha
self.beta = beta
self.gamma = gamma
self.desired_pos = np.array([-0.15, 0.185, 0])
self.objects_initial_positions = {
"ring": [0.53, 0.015, 0],
"ball": [0.4, 0.025, 0],
"box": [0.5, 0.02, 0]
}
self._supervisor = Supervisor()
self._timestep = 32
self._touch = 0
self._num_rew = 0
# List of objects to be taken into consideration
self._objects_names = []
self._objects = {}
# Get Fingers
self.finger_names = ["TouchSensor1", "TouchSensor2"]
self.fingers = {}
for fin in self.finger_names:
self.fingers[fin] = self._supervisor.getFromDef(fin)
# List of links name
self._link_names = [
"arm1", "arm2", "arm3", "arm4", "arm5", "finger1", "finger2"
]
# Get motors
self._link_objects = {}
for link in self._link_names:
self._link_objects[link] = self._supervisor.getMotor(link)
# Get Touch sensor (Added)
self._touch_sensor_names = ["touch sensor1", "touch sensor2"]
#Added
self._touch_sensor_object = {}
for sensor in self._touch_sensor_names:
self._touch_sensor_object[
sensor] = self._supervisor.getTouchSensor(sensor)
self._touch_sensor_object[sensor].enable(self._timestep)
# Get sensors
self._link_position_sensors = {}
self._min_position = {}
self._max_position = {}
for link in self._link_names:
self._link_position_sensors[link] = self._link_objects[
link].getPositionSensor()
self._link_position_sensors[link].enable(self._timestep)
self._min_position[link] = self._link_objects[link].getMinPosition(
)
self._max_position[link] = self._link_objects[link].getMaxPosition(
)
self._supervisor.step(self._timestep)
# self.camera = self._supervisor.getCamera("camera");
# self.camera.enable(self._timestep);
self._floor = 0
self._finger = 0
self._touch = 0
self._i = 0
###### UTIL FUNCTIONS - START ######
def get_links_num(self):
return len(self._link_names)
def set_timestep_simulation(self, step):
self._timestep = step
def set_objects_names(self, names):
self._objects_names = names
for obj in self._objects_names:
self._objects[obj] = self._supervisor.getFromDef(obj)
def set_finger_name(self, name):
self.finger_names = name
self.finger = self._supervisor.getFromDef(name)
def get_link_positions(self):
positions = []
for link in self._link_names:
positions.append(self._link_position_sensors[link].getValue())
return np.array(positions)
def set_link_positions(self, positions):
assert (len(positions) == len(self._link_names))
i = 0
for link in self._link_names:
self._link_objects[link].setPosition(positions[i])
i = i + 1
def get_objects_positions(self):
obj_positions = {}
for obj in self._objects_names:
obj_positions[obj] = self._objects[obj].getPosition()
return obj_positions
def object_position(self):
obj_position = self._objects[self._objects_names[0]].getPosition()
return obj_position
def get_state(self):
state = []
for link in self._link_names:
state = state + [self._link_position_sensors[link].getValue()]
for obj in self._objects_names:
state = state + self._objects[obj].getPosition()
return np.array(state)
def object_finger_distance(self, object_name):
finger_pos_00 = self.fingers["TouchSensor1"].getPosition()
finger_pos_00_array = np.array(finger_pos_00)
finger_pos_01 = self.fingers["TouchSensor2"].getPosition()
finger_pos_01_array = np.array(finger_pos_01)
obj_pos = self.get_objects_positions()[object_name]
obj_pos_array = np.array(obj_pos)
if (object_name == "ring" or object_name == "box"):
finger_distances = [
np.linalg.norm(finger_pos_00_array - obj_pos_array +
[0, 0, +0.019]),
np.linalg.norm(finger_pos_01_array - obj_pos_array +
[0, 0, 0.019])
]
else:
finger_distances = [
np.linalg.norm(finger_pos_00_array - obj_pos_array +
[0, 0, +0]),
np.linalg.norm(finger_pos_01_array - obj_pos_array + [0, 0, 0])
]
return np.array(finger_distances).mean()
def randomizeEnvironment(self):
object_notToGrasp = self._supervisor.getFromDef(self._objects_names[1])
_translation = object_notToGrasp.getField("translation")
array_pos = []
array_pos.append(object_notToGrasp.getPosition())
array_pos.append([1, 0, 1])
array_pos.append([-0.15, 0.195, 0])
rand = randrange(0, 3)
new_position = array_pos[rand]
_translation.setSFVec3f(new_position)
###### UTIL FUNCTIONS - END ######
###### GYM FUNCTIONS - START ######
def step(self, action):
new_state = []
reward = 0
done = False
obs = []
obs = self._get_obs()
reward = self._compute_reward(obs, done)
self.set_link_positions(action)
self._supervisor.step(self._timestep)
return new_state, reward, done, obs
def _compute_reward(self, observations, done):
sensor_mean = (observations["TOUCH_SENSORS"] / 200).mean()
obj_pos = observations["OBJECT_POSITION"]
floor_distance = self.alpha * np.exp(
-(np.linalg.norm(obj_pos - self.desired_pos)**2) /
(2 * dist_dev_alpha**2))
finger_distance = self.beta * np.exp(-(
(self.object_finger_distance(self._objects_names[0]) + 0.9)**2) /
(2 * dist_dev_beta**2))
touch = self.gamma * sensor_mean
touch_distance = touch * np.exp(
-(self.object_finger_distance(self._objects_names[0])**2) /
(2 * dist_dev_gamma**2))
self._num_rew += 1
self._touch += touch_distance
#reward = floor_distance #Per spostamento
reward = finger_distance + touch_distance #Per presa
self._floor += floor_distance
self._finger += finger_distance
self._touch += touch_distance
self._i += 1
'''
print("------------------------")
print("Floor Distance: ")
print(floor_distance)
print("Finger Distance")
print(finger_distance)
print("Reward: ")
print(reward)
print("------------------------")
'''
# print(
# "Dfl: %-10.8f Dfi: %-10.8f D*T: %-10.8f"
# % (floor_distance, finger_distance, touch_distance)
# )
return reward
def _get_obs(self):
joint_positions = self.get_link_positions()
touch_sensors = np.array([
np.linalg.norm(self._touch_sensor_object[sensor].getValues())
for sensor in self._touch_sensor_names
])
obj = self.object_position()
obj_position = np.array(obj)
obs = {
"JOINT_POSITIONS": joint_positions,
"TOUCH_SENSORS": touch_sensors,
"OBJECT_POSITION": obj_position,
}
return obs
def reset(self):
#print("Reset")
#@
if (self._i != 0):
print("floor: ", self._floor, " finger: ", self._finger,
" touch: ", self._touch)
self._i = 0
self._floor = 0
self._finger = 0
self._touch = 0
self._supervisor.simulationReset()
self._supervisor.simulationResetPhysics()
self._supervisor.step(1)
#self.randomizeEnvironment()
for link in self._link_names:
self._link_position_sensors[link].enable(self._timestep)
for sensor in self._touch_sensor_names:
self._touch_sensor_object[sensor].enable(self._timestep)
def render(self, mode='human', close=False):
pass
###### GYM FUNCTIONS - END ######
###### CLASSIFIER FUNCITIONS - START ######
def _objectPositionClassifier(self, object_position,
object_initial_position):
object_position = np.array(object_position)
object_initial_position = np.array(object_initial_position)
diff = np.linalg.norm(object_position - object_initial_position)
return diff <= 0.001
def _fingerDistanceClassifier(self, finger_distance):
return finger_distance <= 0.05
def _desiredPosClassifier(self, object_position, desired_pos):
diff = np.linalg.norm(object_position - desired_pos)
return diff <= 0.03
def _touchSensorsClassifier(self, touch_sensors):
esito = True
for touch_sensor in touch_sensors:
esito = esito and (touch_sensor >= 300)
return esito
def _jointPositionClassifier(self, joint_positions):
esito = True
for joint_position in joint_positions:
esito = esito and (abs(joint_position) <= 0.01)
return esito
def _getValuesFromSensors(self):
obs = self._get_obs()
values = {}
i = 0
obj_names = self._objects_names.copy()
obj_names.sort()
for obj_name in self._objects_names:
values[i] = self._objectPositionClassifier(
self.get_objects_positions()[obj_name],
self.objects_initial_positions[obj_name])
values[i + 1] = self._fingerDistanceClassifier(
self.object_finger_distance(obj_name))
values[i + 2] = self._desiredPosClassifier(
self.get_objects_positions()[obj_name], self.desired_pos)
i += 3
values[i] = self._touchSensorsClassifier(obs["TOUCH_SENSORS"])
values[i + 1] = self._jointPositionClassifier(
self.get_link_positions())
return values