116 lines
4.2 KiB
Python
116 lines
4.2 KiB
Python
import os
|
|
from typing import Optional
|
|
|
|
from gym.envs.mujoco.hopper_v3 import HopperEnv
|
|
import numpy as np
|
|
|
|
MAX_EPISODE_STEPS_HOPPERTHROW = 250
|
|
|
|
|
|
class ALRHopperThrowEnv(HopperEnv):
|
|
"""
|
|
Initialization changes to normal Hopper:
|
|
- healthy_reward: 1.0 -> 0.0 -> 0.1
|
|
- forward_reward_weight -> 5.0
|
|
- healthy_angle_range: (-0.2, 0.2) -> (-float('inf'), float('inf'))
|
|
|
|
Reward changes to normal Hopper:
|
|
- velocity: (x_position_after - x_position_before) -> self.get_body_com("ball")[0]
|
|
"""
|
|
|
|
def __init__(self,
|
|
xml_file='hopper_throw.xml',
|
|
forward_reward_weight=5.0,
|
|
ctrl_cost_weight=1e-3,
|
|
healthy_reward=0.1,
|
|
terminate_when_unhealthy=True,
|
|
healthy_state_range=(-100.0, 100.0),
|
|
healthy_z_range=(0.7, float('inf')),
|
|
healthy_angle_range=(-float('inf'), float('inf')),
|
|
reset_noise_scale=5e-3,
|
|
context=True,
|
|
exclude_current_positions_from_observation=True,
|
|
max_episode_steps=250):
|
|
xml_file = os.path.join(os.path.dirname(__file__), "assets", xml_file)
|
|
self.current_step = 0
|
|
self.max_episode_steps = max_episode_steps
|
|
self.context = context
|
|
self.goal = 0
|
|
super().__init__(xml_file, forward_reward_weight, ctrl_cost_weight, healthy_reward, terminate_when_unhealthy,
|
|
healthy_state_range, healthy_z_range, healthy_angle_range, reset_noise_scale,
|
|
exclude_current_positions_from_observation)
|
|
|
|
def step(self, action):
|
|
self.current_step += 1
|
|
self.do_simulation(action, self.frame_skip)
|
|
ball_pos_after = self.get_body_com("ball")[
|
|
0] # abs(self.get_body_com("ball")[0]) # use x and y to get point and use euclid distance as reward?
|
|
ball_pos_after_y = self.get_body_com("ball")[2]
|
|
|
|
# done = self.done TODO We should use this, not sure why there is no other termination; ball_landed should be enough, because we only look at the throw itself? - Paul and Marc
|
|
ball_landed = bool(self.get_body_com("ball")[2] <= 0.05)
|
|
done = ball_landed
|
|
|
|
ctrl_cost = self.control_cost(action)
|
|
costs = ctrl_cost
|
|
|
|
rewards = 0
|
|
|
|
if self.current_step >= self.max_episode_steps or done:
|
|
distance_reward = -np.linalg.norm(ball_pos_after - self.goal) if self.context else \
|
|
self._forward_reward_weight * ball_pos_after
|
|
healthy_reward = 0 if self.context else self.healthy_reward * self.current_step
|
|
|
|
rewards = distance_reward + healthy_reward
|
|
|
|
observation = self._get_obs()
|
|
reward = rewards - costs
|
|
info = {
|
|
'ball_pos': ball_pos_after,
|
|
'ball_pos_y': ball_pos_after_y,
|
|
'_steps': self.current_step,
|
|
'goal': self.goal,
|
|
}
|
|
|
|
return observation, reward, done, info
|
|
|
|
def _get_obs(self):
|
|
return np.append(super()._get_obs(), self.goal)
|
|
|
|
def reset(self, *, seed: Optional[int] = None, return_info: bool = False, options: Optional[dict] = None):
|
|
self.current_step = 0
|
|
self.goal = self.goal = self.np_random.uniform(2.0, 6.0, 1) # 0.5 8.0
|
|
return super().reset()
|
|
|
|
# overwrite reset_model to make it deterministic
|
|
def reset_model(self):
|
|
noise_low = -self._reset_noise_scale
|
|
noise_high = self._reset_noise_scale
|
|
|
|
qpos = self.init_qpos # + self.np_random.uniform(low=noise_low, high=noise_high, size=self.model.nq)
|
|
qvel = self.init_qvel # + self.np_random.uniform(low=noise_low, high=noise_high, size=self.model.nv)
|
|
|
|
self.set_state(qpos, qvel)
|
|
|
|
observation = self._get_obs()
|
|
return observation
|
|
|
|
|
|
if __name__ == '__main__':
|
|
render_mode = "human" # "human" or "partial" or "final"
|
|
env = ALRHopperThrowEnv()
|
|
obs = env.reset()
|
|
|
|
for i in range(2000):
|
|
# objective.load_result("/tmp/cma")
|
|
# test with random actions
|
|
ac = env.action_space.sample()
|
|
obs, rew, d, info = env.step(ac)
|
|
if i % 10 == 0:
|
|
env.render(mode=render_mode)
|
|
if d:
|
|
print('After ', i, ' steps, done: ', d)
|
|
env.reset()
|
|
|
|
env.close()
|