Improve PyBullet ports of Roboschool envs: fix reset (it kept adding stadium objects, causing slowdown), now reset uses saveState/restoreState and reset becomes a few orders of magnitude faster.
Use python -m pybullet_envs.examples.testEnv --env AntBulletEnv-v0 --render=1 --steps 1000 --resetbenchmark=1 Added environments: HumanoidFlagrunBulletEnv-v0, HumanoidFlagrunHarderBulletEnv-v0, StrikerBulletEnv-v0, ThrowerBulletEnv-v0, PusherBulletEnv-v0, ReacherBulletEnv-v0, CartPoleBulletEnv-v0 and register them to OpenAI Gym. Allow numpy/humanoid_running.py to use abtch or non-batch update (setJointMotorControl2/setJointMotorControlArray)
This commit is contained in:
erwincoumans
234
examples/pybullet/gym/pybullet_envs/gym_manipulator_envs.py
Normal file
234
examples/pybullet/gym/pybullet_envs/gym_manipulator_envs.py
Normal file
@@ -0,0 +1,234 @@
|
||||
from .scene_abstract import SingleRobotEmptyScene
|
||||
from .env_bases import MJCFBaseBulletEnv
|
||||
import numpy as np
|
||||
from robot_manipulators import Reacher, Pusher, Striker, Thrower
|
||||
|
||||
|
||||
class ReacherBulletEnv(MJCFBaseBulletEnv):
|
||||
def __init__(self):
|
||||
self.robot = Reacher()
|
||||
MJCFBaseBulletEnv.__init__(self, self.robot)
|
||||
|
||||
def create_single_player_scene(self):
|
||||
return SingleRobotEmptyScene(gravity=0.0, timestep=0.0165, frame_skip=1)
|
||||
|
||||
def _step(self, a):
|
||||
assert (not self.scene.multiplayer)
|
||||
self.robot.apply_action(a)
|
||||
self.scene.global_step()
|
||||
|
||||
state = self.robot.calc_state() # sets self.to_target_vec
|
||||
|
||||
potential_old = self.potential
|
||||
self.potential = self.robot.calc_potential()
|
||||
|
||||
electricity_cost = (
|
||||
-0.10 * (np.abs(a[0] * self.robot.theta_dot) + np.abs(a[1] * self.robot.gamma_dot)) # work torque*angular_velocity
|
||||
- 0.01 * (np.abs(a[0]) + np.abs(a[1])) # stall torque require some energy
|
||||
)
|
||||
stuck_joint_cost = -0.1 if np.abs(np.abs(self.robot.gamma) - 1) < 0.01 else 0.0
|
||||
self.rewards = [float(self.potential - potential_old), float(electricity_cost), float(stuck_joint_cost)]
|
||||
self.HUD(state, a, False)
|
||||
return state, sum(self.rewards), False, {}
|
||||
|
||||
def camera_adjust(self):
|
||||
x, y, z = self.robot.fingertip.pose().xyz()
|
||||
x *= 0.5
|
||||
y *= 0.5
|
||||
self.camera.move_and_look_at(0.3, 0.3, 0.3, x, y, z)
|
||||
|
||||
|
||||
class PusherBulletEnv(MJCFBaseBulletEnv):
|
||||
def __init__(self):
|
||||
self.robot = Pusher()
|
||||
BaseBulletEnv.__init__(self, self.robot)
|
||||
|
||||
def create_single_player_scene(self):
|
||||
return SingleRobotEmptyScene(gravity=9.81, timestep=0.0020, frame_skip=5)
|
||||
|
||||
def _step(self, a):
|
||||
self.robot.apply_action(a)
|
||||
self.scene.global_step()
|
||||
|
||||
state = self.robot.calc_state() # sets self.to_target_vec
|
||||
|
||||
potential_old = self.potential
|
||||
self.potential = self.robot.calc_potential()
|
||||
|
||||
joint_vel = np.array([
|
||||
self.robot.shoulder_pan_joint.get_velocity(),
|
||||
self.robot.shoulder_lift_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.elbow_flex_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.wrist_flex_joint.get_velocity(),
|
||||
self.robot.wrist_roll_joint.get_velocity()
|
||||
])
|
||||
|
||||
action_product = np.matmul(np.abs(a), np.abs(joint_vel))
|
||||
action_sum = np.sum(a)
|
||||
|
||||
electricity_cost = (
|
||||
-0.10 * action_product # work torque*angular_velocity
|
||||
- 0.01 * action_sum # stall torque require some energy
|
||||
)
|
||||
|
||||
stuck_joint_cost = 0
|
||||
for j in self.robot.ordered_joints:
|
||||
if np.abs(j.current_relative_position()[0]) - 1 < 0.01:
|
||||
stuck_joint_cost += -0.1
|
||||
|
||||
self.rewards = [float(self.potential - potential_old), float(electricity_cost), float(stuck_joint_cost)]
|
||||
self.HUD(state, a, False)
|
||||
return state, sum(self.rewards), False, {}
|
||||
|
||||
def calc_potential(self):
|
||||
return -100 * np.linalg.norm(self.to_target_vec)
|
||||
|
||||
def camera_adjust(self):
|
||||
x, y, z = self.robot.fingertip.pose().xyz()
|
||||
x *= 0.5
|
||||
y *= 0.5
|
||||
self.camera.move_and_look_at(0.3, 0.3, 0.3, x, y, z)
|
||||
|
||||
|
||||
class StrikerBulletEnv(MJCFBaseBulletEnv):
|
||||
def __init__(self):
|
||||
self.robot = Striker()
|
||||
BaseBulletEnv.__init__(self, self.robot)
|
||||
self._striked = False
|
||||
self._min_strike_dist = np.inf
|
||||
self.strike_threshold = 0.1
|
||||
|
||||
def create_single_player_scene(self):
|
||||
return SingleRobotEmptyScene(gravity=9.81, timestep=0.0020, frame_skip=5)
|
||||
|
||||
def _step(self, a):
|
||||
self.robot.apply_action(a)
|
||||
self.scene.global_step()
|
||||
|
||||
state = self.robot.calc_state() # sets self.to_target_vec
|
||||
|
||||
potential_old = self.potential
|
||||
self.potential = self.robot.calc_potential()
|
||||
|
||||
joint_vel = np.array([
|
||||
self.robot.shoulder_pan_joint.get_velocity(),
|
||||
self.robot.shoulder_lift_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.elbow_flex_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.wrist_flex_joint.get_velocity(),
|
||||
self.robot.wrist_roll_joint.get_velocity()
|
||||
])
|
||||
|
||||
action_product = np.matmul(np.abs(a), np.abs(joint_vel))
|
||||
action_sum = np.sum(a)
|
||||
|
||||
electricity_cost = (
|
||||
-0.10 * action_product # work torque*angular_velocity
|
||||
- 0.01 * action_sum # stall torque require some energy
|
||||
)
|
||||
|
||||
stuck_joint_cost = 0
|
||||
for j in self.robot.ordered_joints:
|
||||
if np.abs(j.current_relative_position()[0]) - 1 < 0.01:
|
||||
stuck_joint_cost += -0.1
|
||||
|
||||
dist_object_finger = self.robot.object.pose().xyz() - self.robot.fingertip.pose().xyz()
|
||||
reward_dist_vec = self.robot.object.pose().xyz() - self.robot.target.pose().xyz() # TODO: Should the object and target really belong to the robot? Maybe split this off
|
||||
|
||||
self._min_strike_dist = min(self._min_strike_dist, np.linalg.norm(reward_dist_vec))
|
||||
|
||||
if np.linalg.norm(dist_object_finger) < self.strike_threshold:
|
||||
self._striked = True
|
||||
self._strike_pos = self.robot.fingertip.pose().xyz()
|
||||
|
||||
if self._striked:
|
||||
reward_near_vec = self.robot.object.pose().xyz() - self._strike_pos
|
||||
else:
|
||||
reward_near_vec = self.robot.object.pose().xyz() - self.robot.fingertip.pose().xyz()
|
||||
|
||||
reward_near = - np.linalg.norm(reward_near_vec)
|
||||
|
||||
reward_dist = - np.linalg.norm(self._min_strike_dist)
|
||||
reward_ctrl = - np.square(a).sum()
|
||||
self.rewards = [float(self.potential - potential_old), float(electricity_cost), float(stuck_joint_cost),
|
||||
3 * reward_dist, 0.1 * reward_ctrl, 0.5 * reward_near]
|
||||
self.HUD(state, a, False)
|
||||
return state, sum(self.rewards), False, {}
|
||||
|
||||
def calc_potential(self):
|
||||
return -100 * np.linalg.norm(self.to_target_vec)
|
||||
|
||||
def camera_adjust(self):
|
||||
x, y, z = self.robot.fingertip.pose().xyz()
|
||||
x *= 0.5
|
||||
y *= 0.5
|
||||
self.camera.move_and_look_at(0.3, 0.3, 0.3, x, y, z)
|
||||
|
||||
|
||||
class ThrowerBulletEnv(MJCFBaseBulletEnv):
|
||||
def __init__(self):
|
||||
self.robot = Thrower()
|
||||
BaseBulletEnv.__init__(self, self.robot)
|
||||
|
||||
def create_single_player_scene(self):
|
||||
return SingleRobotEmptyScene(gravity=0.0, timestep=0.0020, frame_skip=5)
|
||||
|
||||
def _step(self, a):
|
||||
self.robot.apply_action(a)
|
||||
self.scene.global_step()
|
||||
state = self.robot.calc_state() # sets self.to_target_vec
|
||||
|
||||
potential_old = self.potential
|
||||
self.potential = self.robot.calc_potential()
|
||||
|
||||
joint_vel = np.array([
|
||||
self.robot.shoulder_pan_joint.get_velocity(),
|
||||
self.robot.shoulder_lift_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.elbow_flex_joint.get_velocity(),
|
||||
self.robot.upper_arm_roll_joint.get_velocity(),
|
||||
self.robot.wrist_flex_joint.get_velocity(),
|
||||
self.robot.wrist_roll_joint.get_velocity()
|
||||
])
|
||||
|
||||
action_product = np.matmul(np.abs(a), np.abs(joint_vel))
|
||||
action_sum = np.sum(a)
|
||||
|
||||
electricity_cost = (
|
||||
-0.10 * action_product # work torque*angular_velocity
|
||||
- 0.01 * action_sum # stall torque require some energy
|
||||
)
|
||||
|
||||
stuck_joint_cost = 0
|
||||
for j in self.robot.ordered_joints:
|
||||
if np.abs(j.current_relative_position()[0]) - 1 < 0.01:
|
||||
stuck_joint_cost += -0.1
|
||||
|
||||
object_xy = self.robot.object.pose().xyz()[:2]
|
||||
target_xy = self.robot.target.pose().xyz()[:2]
|
||||
|
||||
if not self.robot._object_hit_ground and self.robot.object.pose().xyz()[2] < -0.25: # TODO: Should the object and target really belong to the robot? Maybe split this off
|
||||
self.robot._object_hit_ground = True
|
||||
self.robot._object_hit_location = self.robot.object.pose().xyz()
|
||||
|
||||
if self.robot._object_hit_ground:
|
||||
object_hit_xy = self.robot._object_hit_location[:2]
|
||||
reward_dist = -np.linalg.norm(object_hit_xy - target_xy)
|
||||
else:
|
||||
reward_dist = -np.linalg.norm(object_xy - target_xy)
|
||||
reward_ctrl = - np.square(a).sum()
|
||||
|
||||
self.rewards = [float(self.potential - potential_old), float(electricity_cost), float(stuck_joint_cost),
|
||||
reward_dist, 0.002 * reward_ctrl]
|
||||
self.HUD(state, a, False)
|
||||
return state, sum(self.rewards), False, {}
|
||||
|
||||
def camera_adjust(self):
|
||||
x, y, z = self.robot.fingertip.pose().xyz()
|
||||
x *= 0.5
|
||||
y *= 0.5
|
||||
self.camera.move_and_look_at(0.3, 0.3, 0.3, x, y, z)
|
||||
|
||||
Reference in New Issue
Block a user