|
from typing import Any, List, Union, Optional |
|
import time |
|
import gym |
|
import numpy as np |
|
from ding.envs import BaseEnv, BaseEnvTimestep, FrameStackWrapper |
|
from ding.torch_utils import to_ndarray, to_list |
|
from ding.envs.common.common_function import affine_transform |
|
from ding.utils import ENV_REGISTRY |
|
|
|
|
|
@ENV_REGISTRY.register('bipedalwalker') |
|
class BipedalWalkerEnv(BaseEnv): |
|
|
|
def __init__(self, cfg: dict) -> None: |
|
self._cfg = cfg |
|
self._init_flag = False |
|
self._act_scale = cfg.act_scale |
|
self._rew_clip = cfg.rew_clip |
|
if "replay_path" in cfg: |
|
self._replay_path = cfg.replay_path |
|
else: |
|
self._replay_path = None |
|
|
|
def reset(self) -> np.ndarray: |
|
if not self._init_flag: |
|
self._env = gym.make('BipedalWalker-v3') |
|
self._observation_space = self._env.observation_space |
|
self._action_space = self._env.action_space |
|
self._reward_space = gym.spaces.Box( |
|
low=self._env.reward_range[0], high=self._env.reward_range[1], shape=(1, ), dtype=np.float32 |
|
) |
|
self._init_flag = True |
|
if hasattr(self, '_seed') and hasattr(self, '_dynamic_seed') and self._dynamic_seed: |
|
np_seed = 100 * np.random.randint(1, 1000) |
|
self._env.seed(self._seed + np_seed) |
|
elif hasattr(self, '_seed'): |
|
self._env.seed(self._seed) |
|
if self._replay_path is not None: |
|
self._env = gym.wrappers.RecordVideo( |
|
self._env, |
|
video_folder=self._replay_path, |
|
episode_trigger=lambda episode_id: True, |
|
name_prefix='rl-video-{}'.format(id(self)) |
|
) |
|
self._eval_episode_return = 0 |
|
obs = self._env.reset() |
|
obs = to_ndarray(obs).astype(np.float32) |
|
return obs |
|
|
|
def close(self) -> None: |
|
if self._init_flag: |
|
self._env.close() |
|
self._init_flag = False |
|
|
|
def render(self) -> None: |
|
self._env.render() |
|
|
|
def seed(self, seed: int, dynamic_seed: bool = True) -> None: |
|
self._seed = seed |
|
self._dynamic_seed = dynamic_seed |
|
np.random.seed(self._seed) |
|
|
|
def step(self, action: np.ndarray) -> BaseEnvTimestep: |
|
assert isinstance(action, np.ndarray), type(action) |
|
if action.shape == (1, ): |
|
action = action.squeeze() |
|
if self._act_scale: |
|
action = affine_transform(action, min_val=self.action_space.low, max_val=self.action_space.high) |
|
|
|
obs, rew, done, info = self._env.step(action) |
|
self._eval_episode_return += rew |
|
if self._rew_clip: |
|
rew = max(-10, rew) |
|
rew = np.float32(rew) |
|
|
|
if done: |
|
info['eval_episode_return'] = self._eval_episode_return |
|
obs = to_ndarray(obs).astype(np.float32) |
|
rew = to_ndarray([rew]) |
|
return BaseEnvTimestep(obs, rew, done, info) |
|
|
|
def enable_save_replay(self, replay_path: Optional[str] = None) -> None: |
|
if replay_path is None: |
|
replay_path = './video' |
|
self._replay_path = replay_path |
|
|
|
def random_action(self) -> np.ndarray: |
|
random_action = self.action_space.sample() |
|
if isinstance(random_action, np.ndarray): |
|
pass |
|
elif isinstance(random_action, int): |
|
random_action = to_ndarray([random_action], dtype=np.int64) |
|
return random_action |
|
|
|
@property |
|
def observation_space(self) -> gym.spaces.Space: |
|
return self._observation_space |
|
|
|
@property |
|
def action_space(self) -> gym.spaces.Space: |
|
return self._action_space |
|
|
|
@property |
|
def reward_space(self) -> gym.spaces.Space: |
|
return self._reward_space |
|
|
|
def __repr__(self) -> str: |
|
return "DI-engine BipedalWalker Env" |
|
|
