class spotGymEnv(gym.Env):
"""The gym environment for spot.
It simulates the locomotion of spot, a quadruped robot. The state space
include the angles, velocities and torques for all the motors and the action
space is the desired motor angle for each motor. The reward function is based
on how far spot walks in 1000 steps and penalizes the energy
expenditure.
"""
metadata = {
"render.modes": ["human", "rgb_array"],
"video.frames_per_second": 50
}
def __init__(self,
distance_weight=1.0,
rotation_weight=1.0,
energy_weight=0.0005,
shake_weight=0.005,
drift_weight=2.0,
rp_weight=0.1,
rate_weight=0.1,
urdf_root=pybullet_data.getDataPath(),
urdf_version=None,
distance_limit=float("inf"),
observation_noise_stdev=SENSOR_NOISE_STDDEV,
self_collision_enabled=True,
motor_velocity_limit=np.inf,
pd_control_enabled=False,
leg_model_enabled=False,
accurate_motor_model_enabled=False,
remove_default_joint_damping=False,
motor_kp=2.0,
motor_kd=0.03,
control_latency=0.0,
pd_latency=0.0,
torque_control_enabled=False,
motor_overheat_protection=False,
hard_reset=False,
on_rack=False,
render=True,
num_steps_to_log=1000,
action_repeat=1,
control_time_step=None,
env_randomizer=SpotEnvRandomizer(),
forward_reward_cap=float("inf"),
reflection=True,
log_path=None,
desired_velocity=0.5,
desired_rate=0.0,
lateral=False,
draw_foot_path=False,
height_field=False,
height_field_iters=2,
AutoStepper=False,
contacts=True):
"""Initialize the spot gym environment.
Args:
urdf_root: The path to the urdf data folder.
urdf_version: [DEFAULT_URDF_VERSION] are allowable
versions. If None, DEFAULT_URDF_VERSION is used.
distance_weight: The weight of the distance term in the reward.
energy_weight: The weight of the energy term in the reward.
shake_weight: The weight of the vertical shakiness term in the reward.
drift_weight: The weight of the sideways drift term in the reward.
distance_limit: The maximum distance to terminate the episode.
observation_noise_stdev: The standard deviation of observation noise.
self_collision_enabled: Whether to enable self collision in the sim.
motor_velocity_limit: The velocity limit of each motor.
pd_control_enabled: Whether to use PD controller for each motor.
leg_model_enabled: Whether to use a leg motor to reparameterize the action
space.
accurate_motor_model_enabled: Whether to use the accurate DC motor model.
remove_default_joint_damping: Whether to remove the default joint damping.
motor_kp: proportional gain for the accurate motor model.
motor_kd: derivative gain for the accurate motor model.
control_latency: It is the delay in the controller between when an
observation is made at some point, and when that reading is reported
back to the Neural Network.
pd_latency: latency of the PD controller loop. PD calculates PWM based on
the motor angle and velocity. The latency measures the time between when
the motor angle and velocity are observed on the microcontroller and
when the true state happens on the motor. It is typically (0.001-
0.002s).
torque_control_enabled: Whether to use the torque control, if set to
False, pose control will be used.
motor_overheat_protection: Whether to shutdown the motor that has exerted
large torque (OVERHEAT_SHUTDOWN_TORQUE) for an extended amount of time
(OVERHEAT_SHUTDOWN_TIME). See ApplyAction() in spot.py for more
details.
hard_reset: Whether to wipe the simulation and load everything when reset
is called. If set to false, reset just place spot back to start
position and set its pose to initial configuration.
on_rack: Whether to place spot on rack. This is only used to debug
the walking gait. In this mode, spot's base is hanged midair so
that its walking gait is clearer to visualize.
render: Whether to render the simulation.
num_steps_to_log: The max number of control steps in one episode that will
be logged. If the number of steps is more than num_steps_to_log, the
environment will still be running, but only first num_steps_to_log will
be recorded in logging.
action_repeat: The number of simulation steps before actions are applied.
control_time_step: The time step between two successive control signals.
env_randomizer: An instance (or a list) of EnvRandomizer(s). An
EnvRandomizer may randomize the physical property of spot, change
the terrrain during reset(), or add perturbation forces during step().
forward_reward_cap: The maximum value that forward reward is capped at.
Disabled (Inf) by default.
log_path: The path to write out logs. For the details of logging, refer to
spot_logging.proto.
Raises:
ValueError: If the urdf_version is not supported.
"""
# Sense Contacts
self.contacts = contacts
# Enable Auto Stepper State Machine
self.AutoStepper = AutoStepper
# Enable Rough Terrain or Not
self.height_field = height_field
self.draw_foot_path = draw_foot_path
# DRAWING FEET PATH
self.prev_feet_path = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0],
[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]])
# CONTROL METRICS
self.desired_velocity = desired_velocity
self.desired_rate = desired_rate
self.lateral = lateral
# Set up logging.
self._log_path = log_path
# @TODO fix logging
# NUM ITERS
self._time_step = 0.01
self._action_repeat = action_repeat
self._num_bullet_solver_iterations = 300
self.logging = None
if pd_control_enabled or accurate_motor_model_enabled:
self._time_step /= NUM_SUBSTEPS
self._num_bullet_solver_iterations /= NUM_SUBSTEPS
self._action_repeat *= NUM_SUBSTEPS
# PD control needs smaller time step for stability.
if control_time_step is not None:
self.control_time_step = control_time_step
else:
# Get Control Timestep
self.control_time_step = self._time_step * self._action_repeat
# TODO: Fix the value of self._num_bullet_solver_iterations.
self._num_bullet_solver_iterations = int(
NUM_SIMULATION_ITERATION_STEPS / self._action_repeat)
# URDF
self._urdf_root = urdf_root
self._self_collision_enabled = self_collision_enabled
self._motor_velocity_limit = motor_velocity_limit
self._observation = []
self._true_observation = []
self._objectives = []
self._objective_weights = [
distance_weight, energy_weight, drift_weight, shake_weight
]
self._env_step_counter = 0
self._num_steps_to_log = num_steps_to_log
self._is_render = render
self._last_base_position = [0, 0, 0]
self._last_base_orientation = [0, 0, 0, 1]
self._distance_weight = distance_weight
self._rotation_weight = rotation_weight
self._energy_weight = energy_weight
self._drift_weight = drift_weight
self._shake_weight = shake_weight
self._rp_weight = rp_weight
self._rate_weight = rate_weight
self._distance_limit = distance_limit
self._observation_noise_stdev = observation_noise_stdev
self._action_bound = 1
self._pd_control_enabled = pd_control_enabled
self._leg_model_enabled = leg_model_enabled
self._accurate_motor_model_enabled = accurate_motor_model_enabled
self._remove_default_joint_damping = remove_default_joint_damping
self._motor_kp = motor_kp
self._motor_kd = motor_kd
self._torque_control_enabled = torque_control_enabled
self._motor_overheat_protection = motor_overheat_protection
self._on_rack = on_rack
self._cam_dist = 1.0
self._cam_yaw = 0
self._cam_pitch = -30
self._forward_reward_cap = forward_reward_cap
self._hard_reset = True
self._last_frame_time = 0.0
self._control_latency = control_latency
self._pd_latency = pd_latency
self._urdf_version = urdf_version
self._ground_id = None
self._reflection = reflection
self._env_randomizer = env_randomizer
# @TODO fix logging
self._episode_proto = None
if self._is_render:
self._pybullet_client = bullet_client.BulletClient(
connection_mode=pybullet.GUI)
else:
self._pybullet_client = bullet_client.BulletClient()
if self._urdf_version is None:
self._urdf_version = DEFAULT_URDF_VERSION
self._pybullet_client.setPhysicsEngineParameter(enableConeFriction=0)
self.seed()
# Only update after HF has been generated
self.height_field = False
self.reset()
observation_high = (self.spot.GetObservationUpperBound() +
OBSERVATION_EPS)
observation_low = (self.spot.GetObservationLowerBound() -
OBSERVATION_EPS)
action_dim = NUM_MOTORS
action_high = np.array([self._action_bound] * action_dim)
self.action_space = spaces.Box(-action_high, action_high)
self.observation_space = spaces.Box(observation_low, observation_high)
self.viewer = None
self._hard_reset = hard_reset # This assignment need to be after reset()
self.goal_reached = False
# Generate HeightField or not
self.height_field = height_field
self.hf = HeightField()
if self.height_field:
# Do 3x for extra roughness
for i in range(height_field_iters):
self.hf._generate_field(self)
def set_env_randomizer(self, env_randomizer):
self._env_randomizer = env_randomizer
def configure(self, args):
self._args = args
def reset(self,
initial_motor_angles=None,
reset_duration=1.0,
desired_velocity=None,
desired_rate=None):
# Use Autostepper
if self.AutoStepper:
self.StateMachine = BezierStepper(dt=self._time_step)
# Shuffle order of states
self.StateMachine.reshuffle()
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_RENDERING, 0)
if self._hard_reset:
self._pybullet_client.resetSimulation()
self._pybullet_client.setPhysicsEngineParameter(
numSolverIterations=int(self._num_bullet_solver_iterations))
self._pybullet_client.setTimeStep(self._time_step)
self._ground_id = self._pybullet_client.loadURDF("%s/plane.urdf" %
self._urdf_root)
if self._reflection:
self._pybullet_client.changeVisualShape(
self._ground_id, -1, rgbaColor=[1, 1, 1, 0.8])
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_PLANAR_REFLECTION,
self._ground_id)
self._pybullet_client.setGravity(0, 0, -9.81)
acc_motor = self._accurate_motor_model_enabled
motor_protect = self._motor_overheat_protection
if self._urdf_version not in spot_URDF_VERSION_MAP:
raise ValueError("%s is not a supported urdf_version." %
self._urdf_version)
else:
self.spot = (spot_URDF_VERSION_MAP[self._urdf_version](
pybullet_client=self._pybullet_client,
action_repeat=self._action_repeat,
urdf_root=self._urdf_root,
time_step=self._time_step,
self_collision_enabled=self._self_collision_enabled,
motor_velocity_limit=self._motor_velocity_limit,
pd_control_enabled=self._pd_control_enabled,
accurate_motor_model_enabled=acc_motor,
remove_default_joint_damping=self.
_remove_default_joint_damping,
motor_kp=self._motor_kp,
motor_kd=self._motor_kd,
control_latency=self._control_latency,
pd_latency=self._pd_latency,
observation_noise_stdev=self._observation_noise_stdev,
torque_control_enabled=self._torque_control_enabled,
motor_overheat_protection=motor_protect,
on_rack=self._on_rack,
np_random=self.np_random,
contacts=self.contacts))
self.spot.Reset(reload_urdf=False,
default_motor_angles=initial_motor_angles,
reset_time=reset_duration)
if self._env_randomizer is not None:
self._env_randomizer.randomize_env(self)
# Also update heightfield if wr are wholly randomizing
if self.height_field:
self.hf.UpdateHeightField()
if desired_velocity is not None:
self.desired_velocity = desired_velocity
if desired_rate is not None:
self.desired_rate = desired_rate
self._pybullet_client.setPhysicsEngineParameter(enableConeFriction=0)
self._env_step_counter = 0
self._last_base_position = [0, 0, 0]
self._last_base_orientation = [0, 0, 0, 1]
self._objectives = []
self._pybullet_client.resetDebugVisualizerCamera(
self._cam_dist, self._cam_yaw, self._cam_pitch, [0, 0, 0])
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_RENDERING, 1)
return self._get_observation()
def seed(self, seed=None):
self.np_random, seed = seeding.np_random(seed)
return [seed]
def _transform_action_to_motor_command(self, action):
if self._leg_model_enabled:
for i, action_component in enumerate(action):
if not (-self._action_bound - ACTION_EPS <= action_component <=
self._action_bound + ACTION_EPS):
raise ValueError("{}th action {} out of bounds.".format(
i, action_component))
action = self.spot.ConvertFromLegModel(action)
return action
def step(self, action):
"""Step forward the simulation, given the action.
Args:
action: A list of desired motor angles for eight motors.
Returns:
observations: The angles, velocities and torques of all motors.
reward: The reward for the current state-action pair.
done: Whether the episode has ended.
info: A dictionary that stores diagnostic information.
Raises:
ValueError: The action dimension is not the same as the number of motors.
ValueError: The magnitude of actions is out of bounds.
"""
self._last_base_position = self.spot.GetBasePosition()
self._last_base_orientation = self.spot.GetBaseOrientation()
# print("ACTION:")
# print(action)
if self._is_render:
# Sleep, otherwise the computation takes less time than real time,
# which will make the visualization like a fast-forward video.
time_spent = time.time() - self._last_frame_time
self._last_frame_time = time.time()
time_to_sleep = self.control_time_step - time_spent
if time_to_sleep > 0:
time.sleep(time_to_sleep)
base_pos = self.spot.GetBasePosition()
# Keep the previous orientation of the camera set by the user.
[yaw, pitch,
dist] = self._pybullet_client.getDebugVisualizerCamera()[8:11]
self._pybullet_client.resetDebugVisualizerCamera(
dist, yaw, pitch, base_pos)
action = self._transform_action_to_motor_command(action)
self.spot.Step(action)
reward = self._reward()
done = self._termination()
self._env_step_counter += 1
# DRAW FOOT PATH
if self.draw_foot_path:
self.DrawFootPath()
return np.array(self._get_observation()), reward, done, {}
def render(self, mode="rgb_array", close=False):
if mode != "rgb_array":
return np.array([])
base_pos = self.spot.GetBasePosition()
view_matrix = self._pybullet_client.computeViewMatrixFromYawPitchRoll(
cameraTargetPosition=base_pos,
distance=self._cam_dist,
yaw=self._cam_yaw,
pitch=self._cam_pitch,
roll=0,
upAxisIndex=2)
proj_matrix = self._pybullet_client.computeProjectionMatrixFOV(
fov=60,
aspect=float(RENDER_WIDTH) / RENDER_HEIGHT,
nearVal=0.1,
farVal=100.0)
(_, _, px, _, _) = self._pybullet_client.getCameraImage(
width=RENDER_WIDTH,
height=RENDER_HEIGHT,
renderer=self._pybullet_client.ER_BULLET_HARDWARE_OPENGL,
viewMatrix=view_matrix,
projectionMatrix=proj_matrix)
rgb_array = np.array(px)
rgb_array = rgb_array[:, :, :3]
return rgb_array
def DrawFootPath(self):
# Get Foot Positions
FL = self._pybullet_client.getLinkState(self.spot.quadruped,
self.spot._foot_id_list[0])[0]
FR = self._pybullet_client.getLinkState(self.spot.quadruped,
self.spot._foot_id_list[1])[0]
BL = self._pybullet_client.getLinkState(self.spot.quadruped,
self.spot._foot_id_list[2])[0]
BR = self._pybullet_client.getLinkState(self.spot.quadruped,
self.spot._foot_id_list[3])[0]
lifetime = 3.0 # sec
self._pybullet_client.addUserDebugLine(self.prev_feet_path[0],
FL, [1, 0, 0],
lifeTime=lifetime)
self._pybullet_client.addUserDebugLine(self.prev_feet_path[1],
FR, [0, 1, 0],
lifeTime=lifetime)
self._pybullet_client.addUserDebugLine(self.prev_feet_path[2],
BL, [0, 0, 1],
lifeTime=lifetime)
self._pybullet_client.addUserDebugLine(self.prev_feet_path[3],
BR, [1, 1, 0],
lifeTime=lifetime)
self.prev_feet_path[0] = FL
self.prev_feet_path[1] = FR
self.prev_feet_path[2] = BL
self.prev_feet_path[3] = BR
def get_spot_motor_angles(self):
"""Get the spot's motor angles.
Returns:
A numpy array of motor angles.
"""
return np.array(self._observation[
MOTOR_ANGLE_OBSERVATION_INDEX:MOTOR_ANGLE_OBSERVATION_INDEX +
NUM_MOTORS])
def get_spot_motor_velocities(self):
"""Get the spot's motor velocities.
Returns:
A numpy array of motor velocities.
"""
return np.array(self._observation[
MOTOR_VELOCITY_OBSERVATION_INDEX:MOTOR_VELOCITY_OBSERVATION_INDEX +
NUM_MOTORS])
def get_spot_motor_torques(self):
"""Get the spot's motor torques.
Returns:
A numpy array of motor torques.
"""
return np.array(self._observation[
MOTOR_TORQUE_OBSERVATION_INDEX:MOTOR_TORQUE_OBSERVATION_INDEX +
NUM_MOTORS])
def get_spot_base_orientation(self):
"""Get the spot's base orientation, represented by a quaternion.
Returns:
A numpy array of spot's orientation.
"""
return np.array(self._observation[BASE_ORIENTATION_OBSERVATION_INDEX:])
def is_fallen(self):
"""Decide whether spot has fallen.
If the up directions between the base and the world is larger (the dot
product is smaller than 0.85) or the base is very low on the ground
(the height is smaller than 0.13 meter), spot is considered fallen.
Returns:
Boolean value that indicates whether spot has fallen.
"""
orientation = self.spot.GetBaseOrientation()
rot_mat = self._pybullet_client.getMatrixFromQuaternion(orientation)
local_up = rot_mat[6:]
pos = self.spot.GetBasePosition()
# or pos[2] < 0.13
return (np.dot(np.asarray([0, 0, 1]), np.asarray(local_up)) < 0.55)
def _termination(self):
position = self.spot.GetBasePosition()
distance = math.sqrt(position[0]**2 + position[1]**2)
return self.is_fallen() or distance > self._distance_limit
def _reward(self):
""" NOTE: reward now consists of:
roll, pitch at desired 0
acc (y,z) = 0
FORWARD-BACKWARD: rate(x,y,z) = 0
--> HIDDEN REWARD: x(+-) velocity reference, not incl. in obs
SPIN: acc(x) = 0, rate(x,y) = 0, rate (z) = rate reference
Also include drift, energy vanilla rewards
"""
current_base_position = self.spot.GetBasePosition()
# get observation
obs = self._get_observation()
# forward_reward = current_base_position[0] - self._last_base_position[0]
# # POSITIVE FOR FORWARD, NEGATIVE FOR BACKWARD | NOTE: HIDDEN
# GETTING TWIST IN BODY FRAME
pos = self.spot.GetBasePosition()
orn = self.spot.GetBaseOrientation()
roll, pitch, yaw = self._pybullet_client.getEulerFromQuaternion(
[orn[0], orn[1], orn[2], orn[3]])
rpy = LA.RPY(roll, pitch, yaw)
R, _ = LA.TransToRp(rpy)
T_wb = LA.RpToTrans(R, np.array([pos[0], pos[1], pos[2]]))
T_bw = LA.TransInv(T_wb)
Adj_Tbw = LA.Adjoint(T_bw)
Vw = np.concatenate(
(self.spot.prev_ang_twist, self.spot.prev_lin_twist))
Vb = np.dot(Adj_Tbw, Vw)
# New Twist in Body Frame
# POSITIVE FOR FORWARD, NEGATIVE FOR BACKWARD | NOTE: HIDDEN
fwd_speed = -Vb[3] # vx
lat_speed = -Vb[4] # vy
# fwd_speed = self.spot.prev_lin_twist[0]
# lat_speed = self.spot.prev_lin_twist[1]
# print("FORWARD SPEED: {} \t STATE SPEED: {}".format(
# fwd_speed, self.desired_velocity))
# self.desired_velocity = 0.4
# Modification for lateral/fwd rewards
reward_max = 1.0
# FORWARD
if not self.lateral:
# f(x)=-(x-desired))^(2)*((1/desired)^2)+1
# to make sure that at 0vel there is 0 reawrd.
# also squishes allowable tolerance
forward_reward = reward_max * np.exp(
-(fwd_speed - self.desired_velocity)**2 / (0.1))
# LATERAL
else:
forward_reward = reward_max * np.exp(
-(lat_speed - self.desired_velocity)**2 / (0.1))
yaw_rate = obs[4]
rot_reward = reward_max * np.exp(-(yaw_rate - self.desired_rate)**2 /
(0.1))
# Make sure that for forward-policy there is the appropriate rotation penalty
if self.desired_velocity != 0:
self._rotation_weight = self._rate_weight
rot_reward = -abs(obs[4])
elif self.desired_rate != 0:
forward_reward = 0.0
# penalty for nonzero roll, pitch
rp_reward = -(abs(obs[0]) + abs(obs[1]))
# print("ROLL: {} \t PITCH: {}".format(obs[0], obs[1]))
# penalty for nonzero acc(z)
shake_reward = -abs(obs[4])
# penalty for nonzero rate (x,y,z)
rate_reward = -(abs(obs[2]) + abs(obs[3]))
# drift_reward = -abs(current_base_position[1] -
# self._last_base_position[1])
# this penalizes absolute error, and does not penalize correction
# NOTE: for side-side, drift reward becomes in x instead
drift_reward = -abs(current_base_position[1])
# If Lateral, change drift reward
if self.lateral:
drift_reward = -abs(current_base_position[0])
# shake_reward = -abs(current_base_position[2] -
# self._last_base_position[2])
self._last_base_position = current_base_position
energy_reward = -np.abs(
np.dot(self.spot.GetMotorTorques(),
self.spot.GetMotorVelocities())) * self._time_step
reward = (self._distance_weight * forward_reward +
self._rotation_weight * rot_reward +
self._energy_weight * energy_reward +
self._drift_weight * drift_reward +
self._shake_weight * shake_reward +
self._rp_weight * rp_reward +
self._rate_weight * rate_reward)
self._objectives.append(
[forward_reward, energy_reward, drift_reward, shake_reward])
# print("REWARD: ", reward)
return reward
def get_objectives(self):
return self._objectives
@property
def objective_weights(self):
"""Accessor for the weights for all the objectives.
Returns:
List of floating points that corresponds to weights for the objectives in
the order that objectives are stored.
"""
return self._objective_weights
def _get_observation(self):
"""Get observation of this environment, including noise and latency.
spot class maintains a history of true observations. Based on the
latency, this function will find the observation at the right time,
interpolate if necessary. Then Gaussian noise is added to this observation
based on self.observation_noise_stdev.
Returns:
The noisy observation with latency.
"""
self._observation = self.spot.GetObservation()
return self._observation
def _get_realistic_observation(self):
"""Get the observations of this environment.
It includes the angles, velocities, torques and the orientation of the base.
Returns:
The observation list. observation[0:8] are motor angles. observation[8:16]
are motor velocities, observation[16:24] are motor torques.
observation[24:28] is the orientation of the base, in quaternion form.
"""
self._observation = self.spot.RealisticObservation()
return self._observation
if parse_version(gym.__version__) < parse_version('0.9.6'):
_render = render
_reset = reset
_seed = seed
_step = step
def set_time_step(self, control_step, simulation_step=0.001):
"""Sets the time step of the environment.
Args:
control_step: The time period (in seconds) between two adjacent control
actions are applied.
simulation_step: The simulation time step in PyBullet. By default, the
simulation step is 0.001s, which is a good trade-off between simulation
speed and accuracy.
Raises:
ValueError: If the control step is smaller than the simulation step.
"""
if control_step < simulation_step:
raise ValueError(
"Control step should be larger than or equal to simulation step."
)
self.control_time_step = control_step
self._time_step = simulation_step
self._action_repeat = int(round(control_step / simulation_step))
self._num_bullet_solver_iterations = (NUM_SIMULATION_ITERATION_STEPS /
self._action_repeat)
self._pybullet_client.setPhysicsEngineParameter(
numSolverIterations=self._num_bullet_solver_iterations)
self._pybullet_client.setTimeStep(self._time_step)
self.spot.SetTimeSteps(action_repeat=self._action_repeat,
simulation_step=self._time_step)
@property
def pybullet_client(self):
return self._pybullet_client
@property
def ground_id(self):
return self._ground_id
@ground_id.setter
def ground_id(self, new_ground_id):
self._ground_id = new_ground_id
@property
def env_step_counter(self):
return self._env_step_counter
def main():
""" The main() function. """
print("STARTING SPOT TEST ENV")
seed = 0
max_timesteps = 4e6
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
models_path = os.path.join(my_path, "../models")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
if ARGS.DebugRack:
on_rack = True
else:
on_rack = False
if ARGS.DebugPath:
draw_foot_path = True
else:
draw_foot_path = False
if ARGS.HeightField:
height_field = True
else:
height_field = False
env = spotBezierEnv(render=True,
on_rack=on_rack,
height_field=height_field,
draw_foot_path=draw_foot_path)
# Set seeds
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
state = env.reset()
g_u_i = GUI(env.spot.quadruped)
spot = SpotModel()
T_bf0 = spot.WorldToFoot
T_bf = copy.deepcopy(T_bf0)
bzg = BezierGait(dt=env._time_step)
bz_step = BezierStepper(dt=env._time_step, mode=0)
action = env.action_space.sample()
FL_phases = []
FR_phases = []
BL_phases = []
BR_phases = []
yaw = 0.0
print("STARTED SPOT TEST ENV")
t = 0
while t < (int(max_timesteps)):
bz_step.ramp_up()
pos, orn, StepLength, LateralFraction, YawRate, StepVelocity, ClearanceHeight, PenetrationDepth = bz_step.StateMachine(
)
pos, orn, StepLength, LateralFraction, YawRate, StepVelocity, ClearanceHeight, PenetrationDepth = g_u_i.UserInput(
)
yaw = env.return_yaw()
P_yaw = 5.0
if ARGS.AutoYaw:
YawRate += -yaw * P_yaw
# print("YAW RATE: {}".format(YawRate))
# TEMP
bz_step.StepLength = StepLength
bz_step.LateralFraction = LateralFraction
bz_step.YawRate = YawRate
bz_step.StepVelocity = StepVelocity
contacts = state[-4:]
FL_phases.append(env.spot.LegPhases[0])
FR_phases.append(env.spot.LegPhases[1])
BL_phases.append(env.spot.LegPhases[2])
BR_phases.append(env.spot.LegPhases[3])
# Get Desired Foot Poses
T_bf = bzg.GenerateTrajectory(StepLength, LateralFraction, YawRate,
StepVelocity, T_bf0, T_bf,
ClearanceHeight, PenetrationDepth,
contacts)
joint_angles = spot.IK(orn, pos, T_bf)
env.pass_joint_angles(joint_angles.reshape(-1))
# Get External Observations
env.spot.GetExternalObservations(bzg, bz_step)
# Step
state, reward, done, _ = env.step(action)
if done:
print("DONE")
if ARGS.AutoReset:
env.reset()
# plt.plot()
# plt.plot(FL_phases, label="FL")
# plt.plot(FR_phases, label="FR")
# plt.plot(BL_phases, label="BL")
# plt.plot(BR_phases, label="BR")
# plt.xlabel("dt")
# plt.ylabel("value")
# plt.title("Leg Phases")
# plt.legend()
# plt.show()
# time.sleep(1.0)
t += 1
env.close()
print(joint_angles)
def reset(self,
initial_motor_angles=None,
reset_duration=1.0,
desired_velocity=None,
desired_rate=None):
# Use Autostepper
if self.AutoStepper:
self.StateMachine = BezierStepper(dt=self._time_step)
# Shuffle order of states
self.StateMachine.reshuffle()
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_RENDERING, 0)
if self._hard_reset:
self._pybullet_client.resetSimulation()
self._pybullet_client.setPhysicsEngineParameter(
numSolverIterations=int(self._num_bullet_solver_iterations))
self._pybullet_client.setTimeStep(self._time_step)
self._ground_id = self._pybullet_client.loadURDF("%s/plane.urdf" %
self._urdf_root)
if self._reflection:
self._pybullet_client.changeVisualShape(
self._ground_id, -1, rgbaColor=[1, 1, 1, 0.8])
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_PLANAR_REFLECTION,
self._ground_id)
self._pybullet_client.setGravity(0, 0, -9.81)
acc_motor = self._accurate_motor_model_enabled
motor_protect = self._motor_overheat_protection
if self._urdf_version not in spot_URDF_VERSION_MAP:
raise ValueError("%s is not a supported urdf_version." %
self._urdf_version)
else:
self.spot = (spot_URDF_VERSION_MAP[self._urdf_version](
pybullet_client=self._pybullet_client,
action_repeat=self._action_repeat,
urdf_root=self._urdf_root,
time_step=self._time_step,
self_collision_enabled=self._self_collision_enabled,
motor_velocity_limit=self._motor_velocity_limit,
pd_control_enabled=self._pd_control_enabled,
accurate_motor_model_enabled=acc_motor,
remove_default_joint_damping=self.
_remove_default_joint_damping,
motor_kp=self._motor_kp,
motor_kd=self._motor_kd,
control_latency=self._control_latency,
pd_latency=self._pd_latency,
observation_noise_stdev=self._observation_noise_stdev,
torque_control_enabled=self._torque_control_enabled,
motor_overheat_protection=motor_protect,
on_rack=self._on_rack,
np_random=self.np_random,
contacts=self.contacts))
self.spot.Reset(reload_urdf=False,
default_motor_angles=initial_motor_angles,
reset_time=reset_duration)
if self._env_randomizer is not None:
self._env_randomizer.randomize_env(self)
# Also update heightfield if wr are wholly randomizing
if self.height_field:
self.hf.UpdateHeightField()
if desired_velocity is not None:
self.desired_velocity = desired_velocity
if desired_rate is not None:
self.desired_rate = desired_rate
self._pybullet_client.setPhysicsEngineParameter(enableConeFriction=0)
self._env_step_counter = 0
self._last_base_position = [0, 0, 0]
self._last_base_orientation = [0, 0, 0, 1]
self._objectives = []
self._pybullet_client.resetDebugVisualizerCamera(
self._cam_dist, self._cam_yaw, self._cam_pitch, [0, 0, 0])
self._pybullet_client.configureDebugVisualizer(
self._pybullet_client.COV_ENABLE_RENDERING, 1)
return self._get_observation()
def main():
""" The main() function. """
print("STARTING MINITAUR ARS")
# TRAINING PARAMETERS
# env_name = "MinitaurBulletEnv-v0"
seed = 0
max_timesteps = 4e6
file_name = "spot_ars_"
if ARGS.DebugRack:
on_rack = True
else:
on_rack = False
if ARGS.DebugPath:
draw_foot_path = True
else:
draw_foot_path = False
if ARGS.HeightField:
height_field = True
else:
height_field = False
if ARGS.NoContactSensing:
contacts = False
else:
contacts = True
if ARGS.DontRender:
render = False
else:
render = True
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
if contacts:
models_path = os.path.join(my_path, "../models/contact")
else:
models_path = os.path.join(my_path, "../models/no_contact")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
env = spotBezierEnv(render=render,
on_rack=on_rack,
height_field=height_field,
draw_foot_path=draw_foot_path,
contacts=contacts)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
spot = SpotModel()
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim)
# to GUI or not to GUI
if ARGS.GUI:
gui = True
else:
gui = False
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot, gui)
agent_num = 0
if ARGS.AgentNum:
agent_num = ARGS.AgentNum
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent")
agent.load(models_path + "/" + file_name + str(agent_num))
agent.policy.episode_steps = np.inf
policy = agent.policy
# Evaluate untrained agent and init list for storage
evaluations = []
env.reset()
episode_reward = 0
episode_timesteps = 0
episode_num = 0
print("STARTED MINITAUR TEST SCRIPT")
t = 0
while t < (int(max_timesteps)):
episode_reward, episode_timesteps = agent.deployTG()
t += episode_timesteps
# episode_reward = agent.train()
# +1 to account for 0 indexing.
# +0 on ep_timesteps since it will increment +1 even if done=True
print("Total T: {} Episode Num: {} Episode T: {} Reward: {}".format(
t, episode_num, episode_timesteps, episode_reward))
episode_num += 1
env.close()
def main():
""" The main() function. """
print("STARTING MINITAUR ARS")
# TRAINING PARAMETERS
# env_name = "MinitaurBulletEnv-v0"
seed = 0
if ARGS.Seed:
seed = ARGS.Seed
max_timesteps = 4e6
file_name = "spot_ars_"
if ARGS.DebugRack:
on_rack = True
else:
on_rack = False
if ARGS.DebugPath:
draw_foot_path = True
else:
draw_foot_path = False
if ARGS.HeightField:
height_field = True
else:
height_field = False
if ARGS.NoContactSensing:
contacts = False
else:
contacts = True
if ARGS.DontRender:
render = False
else:
render = True
if ARGS.DontRandomize:
env_randomizer = None
else:
env_randomizer = SpotEnvRandomizer()
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
if contacts:
models_path = os.path.join(my_path, "../models/contact")
else:
models_path = os.path.join(my_path, "../models/no_contact")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
env = spotBezierEnv(render=render,
on_rack=on_rack,
height_field=height_field,
draw_foot_path=draw_foot_path,
contacts=contacts,
env_randomizer=env_randomizer)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
spot = SpotModel()
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim)
# to GUI or not to GUI
if ARGS.GUI:
gui = True
else:
gui = False
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot, gui)
agent_num = 0
if ARGS.AgentNum:
agent_num = ARGS.AgentNum
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent")
agent.load(models_path + "/" + file_name + str(agent_num))
agent.policy.episode_steps = np.inf
policy = agent.policy
env.reset()
episode_reward = 0
episode_timesteps = 0
episode_num = 0
print("STARTED MINITAUR TEST SCRIPT")
t = 0
while t < (int(max_timesteps)):
episode_reward, episode_timesteps = agent.deployTG()
t += episode_timesteps
# episode_reward = agent.train()
# +1 to account for 0 indexing.
# +0 on ep_timesteps since it will increment +1 even if done=True
print("Total T: {} Episode Num: {} Episode T: {} Reward: {}".format(
t, episode_num, episode_timesteps, episode_reward))
episode_num += 1
# Plot Policy Output
if ARGS.PlotPolicy or ARGS.TrueAction or ARGS.SaveData:
if ARGS.TrueAction:
action_name = "robot_act"
action = np.array(agent.true_action_history)
else:
action_name = "agent_act"
action = np.array(agent.action_history)
if ARGS.SaveData:
if height_field:
terrain_name = "rough_"
else:
terrain_name = "flat_"
np.save(
results_path + "/" + "policy_out_" + terrain_name +
action_name, action)
print("SAVED DATA")
ClearHeight_act = action[:, 0]
BodyHeight_act = action[:, 1]
Residuals_act = action[:, 2:]
plt.plot(ClearHeight_act,
label='Clearance Height Mod',
color='black')
plt.plot(BodyHeight_act,
label='Body Height Mod',
color='darkviolet')
# FL
plt.plot(Residuals_act[:, 0],
label='Residual: FL (x)',
color='limegreen')
plt.plot(Residuals_act[:, 1],
label='Residual: FL (y)',
color='lime')
plt.plot(Residuals_act[:, 2],
label='Residual: FL (z)',
color='green')
# FR
plt.plot(Residuals_act[:, 3],
label='Residual: FR (x)',
color='lightskyblue')
plt.plot(Residuals_act[:, 4],
label='Residual: FR (y)',
color='dodgerblue')
plt.plot(Residuals_act[:, 5],
label='Residual: FR (z)',
color='blue')
# BL
plt.plot(Residuals_act[:, 6],
label='Residual: BL (x)',
color='firebrick')
plt.plot(Residuals_act[:, 7],
label='Residual: BL (y)',
color='crimson')
plt.plot(Residuals_act[:, 8],
label='Residual: BL (z)',
color='red')
# BR
plt.plot(Residuals_act[:, 9],
label='Residual: BR (x)',
color='gold')
plt.plot(Residuals_act[:, 10],
label='Residual: BR (y)',
color='orange')
plt.plot(Residuals_act[:, 11],
label='Residual: BR (z)',
color='coral')
plt.xlabel("Epoch Iteration")
plt.ylabel("Action Value")
plt.title("Policy Output")
plt.legend()
plt.show()
env.close()
def main():
""" The main() function. """
# Hold mp pipes
mp.freeze_support()
print("STARTING SPOT TRAINING ENV")
seed = 0
max_timesteps = 4e6
eval_freq = 1e1
save_model = True
file_name = "spot_ars_"
if ARGS.HeightField:
height_field = True
else:
height_field = False
if ARGS.NoContactSensing:
contacts = False
else:
contacts = True
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
if contacts:
models_path = os.path.join(my_path, "../models/contact")
else:
models_path = os.path.join(my_path, "../models/no_contact")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
env = spotBezierEnv(render=False,
on_rack=False,
height_field=height_field,
draw_foot_path=False,
contacts=contacts)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
g_u_i = GUI(env.spot.quadruped)
spot = SpotModel()
T_bf = spot.WorldToFoot
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim)
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot)
agent_num = 0
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent")
agent.load(models_path + "/" + file_name + str(agent_num))
env.reset(agent.desired_velocity, agent.desired_rate)
episode_reward = 0
episode_timesteps = 0
episode_num = 0
# Create mp pipes
num_processes = policy.num_deltas
processes = []
childPipes = []
parentPipes = []
# Store mp pipes
for pr in range(num_processes):
parentPipe, childPipe = Pipe()
parentPipes.append(parentPipe)
childPipes.append(childPipe)
# Start multiprocessing
# Start multiprocessing
for proc_num in range(num_processes):
p = mp.Process(target=ParallelWorker,
args=(childPipes[proc_num], env, state_dim))
p.start()
processes.append(p)
print("STARTED SPOT TRAINING ENV")
t = 0
while t < (int(max_timesteps)):
# Maximum timesteps per rollout
episode_reward, episode_timesteps = agent.train_parallel(parentPipes)
t += episode_timesteps
# episode_reward = agent.train()
# +1 to account for 0 indexing.
# +0 on ep_timesteps since it will increment +1 even if done=True
print(
"Total T: {} Episode Num: {} Episode T: {} Reward: {:.2f} REWARD PER STEP: {:.2f}"
.format(t + 1, episode_num, episode_timesteps, episode_reward,
episode_reward / float(episode_timesteps)))
# Evaluate episode
if (episode_num + 1) % eval_freq == 0:
if save_model:
agent.save(models_path + "/" + str(file_name) +
str(episode_num))
# replay_buffer.save(t)
episode_num += 1
# Close pipes and hence envs
for parentPipe in parentPipes:
parentPipe.send([_CLOSE, "pay2"])
for p in processes:
p.join()
def main():
""" The main() function. """
print("STARTING SPOT SAC")
# TRAINING PARAMETERS
seed = 0
max_timesteps = 4e6
batch_size = 256
eval_freq = 1e4
save_model = True
file_name = "spot_sac_"
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
models_path = os.path.join(my_path, "../models")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
env = spotBezierEnv(render=False,
on_rack=False,
height_field=False,
draw_foot_path=False)
env = NormalizedActions(env)
# Set seeds
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
print("RECORDED MAX ACTION: {}".format(max_action))
hidden_dim = 256
policy = PolicyNetwork(state_dim, action_dim, hidden_dim)
replay_buffer_size = 1000000
replay_buffer = ReplayBuffer(replay_buffer_size)
sac = SoftActorCritic(policy=policy,
state_dim=state_dim,
action_dim=action_dim,
replay_buffer=replay_buffer)
policy_num = 0
if os.path.exists(models_path + "/" + file_name + str(policy_num) +
"_critic"):
print("Loading Existing Policy")
sac.load(models_path + "/" + file_name + str(policy_num))
policy = sac.policy_net
# Evaluate untrained policy and init list for storage
evaluations = []
state = env.reset()
done = False
episode_reward = 0
episode_timesteps = 0
episode_num = 0
max_t_per_ep = 5000
# State Machine for Random Controller Commands
bz_step = BezierStepper(dt=0.01)
# Bezier Gait Generator
bzg = BezierGait(dt=0.01)
# Spot Model
spot = SpotModel()
T_bf0 = spot.WorldToFoot
T_bf = copy.deepcopy(T_bf0)
BaseClearanceHeight = bz_step.ClearanceHeight
BasePenetrationDepth = bz_step.PenetrationDepth
print("STARTED SPOT SAC")
for t in range(int(max_timesteps)):
pos, orn, StepLength, LateralFraction, YawRate, StepVelocity, ClearanceHeight, PenetrationDepth = bz_step.StateMachine(
)
env.spot.GetExternalObservations(bzg, bz_step)
# Read UPDATED state based on controls and phase
state = env.return_state()
action = sac.policy_net.get_action(state)
# Bezier params specced by action
CD_SCALE = 0.002
SLV_SCALE = 0.01
StepLength += action[0] * CD_SCALE
StepVelocity += action[1] * SLV_SCALE
LateralFraction += action[2] * SLV_SCALE
YawRate = action[3]
ClearanceHeight += action[4] * CD_SCALE
PenetrationDepth += action[5] * CD_SCALE
# CLIP EVERYTHING
StepLength = np.clip(StepLength, bz_step.StepLength_LIMITS[0],
bz_step.StepLength_LIMITS[1])
StepVelocity = np.clip(StepVelocity, bz_step.StepVelocity_LIMITS[0],
bz_step.StepVelocity_LIMITS[1])
LateralFraction = np.clip(LateralFraction,
bz_step.LateralFraction_LIMITS[0],
bz_step.LateralFraction_LIMITS[1])
YawRate = np.clip(YawRate, bz_step.YawRate_LIMITS[0],
bz_step.YawRate_LIMITS[1])
ClearanceHeight = np.clip(ClearanceHeight,
bz_step.ClearanceHeight_LIMITS[0],
bz_step.ClearanceHeight_LIMITS[1])
PenetrationDepth = np.clip(PenetrationDepth,
bz_step.PenetrationDepth_LIMITS[0],
bz_step.PenetrationDepth_LIMITS[1])
contacts = state[-4:]
# Get Desired Foot Poses
T_bf = bzg.GenerateTrajectory(StepLength, LateralFraction, YawRate,
StepVelocity, T_bf0, T_bf,
ClearanceHeight, PenetrationDepth,
contacts)
# Add DELTA to XYZ Foot Poses
RESIDUALS_SCALE = 0.05
# T_bf["FL"][3, :3] += action[6:9] * RESIDUALS_SCALE
# T_bf["FR"][3, :3] += action[9:12] * RESIDUALS_SCALE
# T_bf["BL"][3, :3] += action[12:15] * RESIDUALS_SCALE
# T_bf["BR"][3, :3] += action[15:18] * RESIDUALS_SCALE
T_bf["FL"][3, 2] += action[6] * RESIDUALS_SCALE
T_bf["FR"][3, 2] += action[7] * RESIDUALS_SCALE
T_bf["BL"][3, 2] += action[8] * RESIDUALS_SCALE
T_bf["BR"][3, 2] += action[9] * RESIDUALS_SCALE
joint_angles = spot.IK(orn, pos, T_bf)
# Pass Joint Angles
env.pass_joint_angles(joint_angles.reshape(-1))
# Perform action
next_state, reward, done, _ = env.step(action)
done_bool = float(done)
episode_timesteps += 1
# Store data in replay buffer
replay_buffer.push(state, action, reward, next_state, done_bool)
state = next_state
episode_reward += reward
# Train agent after collecting sufficient data for buffer
if len(replay_buffer) > batch_size:
sac.soft_q_update(batch_size)
if episode_timesteps > max_t_per_ep:
done = True
if done:
# Reshuffle State Machine
bzg.reset()
bz_step.reshuffle()
bz_step.ClearanceHeight = BaseClearanceHeight
bz_step.PenetrationDepth = BasePenetrationDepth
# +1 to account for 0 indexing.
# +0 on ep_timesteps since it will increment +1 even if done=True
print(
"Total T: {} Episode Num: {} Episode T: {} Reward: {:.2f} REWARD PER STEP: {:.2f}"
.format(t + 1, episode_num, episode_timesteps, episode_reward,
episode_reward / float(episode_timesteps)))
# Reset environment
state, done = env.reset(), False
evaluations.append(episode_reward)
episode_reward = 0
episode_timesteps = 0
episode_num += 1
# Evaluate episode
if (t + 1) % eval_freq == 0:
# evaluate_policy(policy, env_name, seed,
np.save(results_path + "/" + str(file_name), evaluations)
if save_model:
sac.save(models_path + "/" + str(file_name) + str(t))
# replay_buffer.save(t)
env.close()
def main():
""" The main() function. """
print("STARTING MINITAUR ARS")
# TRAINING PARAMETERS
# env_name = "MinitaurBulletEnv-v0"
seed = 0
max_timesteps = 4e6
file_name = "spot_ars_"
if ARGS.DebugRack:
on_rack = True
else:
on_rack = False
if ARGS.DebugPath:
draw_foot_path = True
else:
draw_foot_path = False
if ARGS.HeightField:
height_field = True
else:
height_field = False
if ARGS.NoContactSensing:
contacts = False
else:
contacts = True
if ARGS.DontRender:
render = False
else:
render = True
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
if contacts:
models_path = os.path.join(my_path, "../models/contact")
else:
models_path = os.path.join(my_path, "../models/no_contact")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
env = spotBezierEnv(render=render,
on_rack=on_rack,
height_field=height_field,
draw_foot_path=draw_foot_path,
contacts=contacts)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
spot = SpotModel()
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim)
# to GUI or not to GUI
if ARGS.GUI:
gui = True
else:
gui = False
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot, gui)
agent_num = 9
still_going = True
print("Loading and Saving")
while still_going:
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent: {}".format(agent_num))
# Load Class
agent.load(models_path + "/" + file_name + str(agent_num))
# Save np array
agent.save(models_path + "/" + file_name + str(agent_num))
else:
still_going = False
agent_num += 10
def main():
""" The main() function. """
print("STARTING MINITAUR ARS")
# TRAINING PARAMETERS
# env_name = "MinitaurBulletEnv-v0"
seed = 0
max_episodes = 1000
if ARGS.NumberOfEpisodes:
max_episodes = ARGS.NumberOfEpisodes
file_name = "spot_ars_"
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
models_path = os.path.join(my_path, "../models")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
if ARGS.HeightField:
height_field = True
else:
height_field = False
env = spotBezierEnv(render=False,
on_rack=False,
height_field=height_field,
draw_foot_path=False)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
spot = SpotModel()
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim, episode_steps=np.inf)
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot, False)
use_agent = False
agent_num = 0
if ARGS.AgentNum:
agent_num = ARGS.AgentNum
use_agent = True
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent")
agent.load(models_path + "/" + file_name + str(agent_num))
agent.policy.episode_steps = 50000
policy = agent.policy
env.reset()
episode_reward = 0
episode_timesteps = 0
episode_num = 0
print("STARTED MINITAUR TEST SCRIPT")
# Used to create gaussian distribution of survival
surv_dt = []
while episode_num < (int(max_episodes)):
episode_reward, episode_timesteps = agent.deployTG()
episode_num += 1
# Store dt and frequency for prob distribution
surv_dt.append(episode_timesteps)
print("Episode Num: {} Episode T: {} Reward: {}".format(
episode_num, episode_timesteps, episode_reward))
env.close()
print("---------------------------------------")
# Store results
if use_agent:
# Store _agent
agt = "agent"
else:
# Store _vanilla
agt = "vanilla"
with open(
results_path + "/" + str(file_name) + agt + '_survival_' +
str(max_episodes), 'wb') as filehandle:
pickle.dump(surv_dt, filehandle)
def main():
""" The main() function. """
print("STARTING MINITAUR ARS")
# TRAINING PARAMETERS
# env_name = "MinitaurBulletEnv-v0"
seed = 0
max_episodes = 1000
if ARGS.NumberOfEpisodes:
max_episodes = ARGS.NumberOfEpisodes
if ARGS.HeightField:
height_field = True
else:
height_field = False
if ARGS.NoContactSensing:
contacts = False
else:
contacts = True
if ARGS.DontRandomize:
env_randomizer = None
else:
env_randomizer = SpotEnvRandomizer()
file_name = "spot_ars_"
# Find abs path to this file
my_path = os.path.abspath(os.path.dirname(__file__))
results_path = os.path.join(my_path, "../results")
if contacts:
models_path = os.path.join(my_path, "../models/contact")
else:
models_path = os.path.join(my_path, "../models/no_contact")
if not os.path.exists(results_path):
os.makedirs(results_path)
if not os.path.exists(models_path):
os.makedirs(models_path)
if ARGS.HeightField:
height_field = True
else:
height_field = False
env = spotBezierEnv(render=False,
on_rack=False,
height_field=height_field,
draw_foot_path=False,
contacts=contacts,
env_randomizer=env_randomizer)
# Set seeds
env.seed(seed)
np.random.seed(seed)
state_dim = env.observation_space.shape[0]
print("STATE DIM: {}".format(state_dim))
action_dim = env.action_space.shape[0]
print("ACTION DIM: {}".format(action_dim))
max_action = float(env.action_space.high[0])
env.reset()
spot = SpotModel()
bz_step = BezierStepper(dt=env._time_step)
bzg = BezierGait(dt=env._time_step)
# Initialize Normalizer
normalizer = Normalizer(state_dim)
# Initialize Policy
policy = Policy(state_dim, action_dim, episode_steps=np.inf)
# Initialize Agent with normalizer, policy and gym env
agent = ARSAgent(normalizer, policy, env, bz_step, bzg, spot, False)
use_agent = False
agent_num = 0
if ARGS.AgentNum:
agent_num = ARGS.AgentNum
use_agent = True
if os.path.exists(models_path + "/" + file_name + str(agent_num) +
"_policy"):
print("Loading Existing agent")
agent.load(models_path + "/" + file_name + str(agent_num))
agent.policy.episode_steps = 50000
policy = agent.policy
env.reset()
episode_reward = 0
episode_timesteps = 0
episode_num = 0
print("STARTED MINITAUR TEST SCRIPT")
# Used to create gaussian distribution of survival distance
surv_pos = []
# Reset every 200 episodes (pb client doesn't like running for long)
reset_ep = 200
while episode_num < (int(max_episodes)):
episode_reward, episode_timesteps = agent.deployTG()
# We only care about x/y pos
travelled_pos = list(agent.returnPose())
# NOTE: FORMAT: X, Y, TIMESTEPS -
# tells us if robobt was just stuck forever. didn't actually fall.
travelled_pos[-1] = episode_timesteps
episode_num += 1
# Store dt and frequency for prob distribution
surv_pos.append(travelled_pos)
print("Episode Num: {} Episode T: {} Reward: {}".format(
episode_num, episode_timesteps, episode_reward))
print("Survival Pos: {}".format(surv_pos[-1]))
# Reset every X episodes (pb client doesn't like running for long)
if episode_num % reset_ep == 0:
env.close()
env = spotBezierEnv(render=False,
on_rack=False,
height_field=height_field,
draw_foot_path=False,
contacts=contacts,
env_randomizer=env_randomizer)
# Set seeds
env.seed(seed)
agent.env = env
env.close()
print("---------------------------------------")
# Store results
if use_agent:
# Store _agent
agt = "agent_" + str(agent_num)
else:
# Store _vanilla
agt = "vanilla"
with open(
results_path + "/" + str(file_name) + agt + '_survival_' +
str(max_episodes), 'wb') as filehandle:
pickle.dump(surv_pos, filehandle)
