1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138 | from __future__ import annotations
from typing import SupportsFloat
import gymnasium as gym
from gymnasium import spaces
import numpy as np
class HistoryWrapper(gym.Wrapper[np.ndarray, np.ndarray, np.ndarray, np.ndarray]):
"""Stack past observations and actions to give a history to the agent.
:param env: Gym environment.
:param horizon: Number of past steps to include in the observation.
"""
def __init__(self, env: gym.Env, horizon: int = 2) -> None:
assert isinstance(env.observation_space, spaces.Box)
assert isinstance(env.action_space, spaces.Box)
wrapped_obs_space = env.observation_space
wrapped_action_space = env.action_space
low_obs = np.tile(wrapped_obs_space.low, horizon)
high_obs = np.tile(wrapped_obs_space.high, horizon)
low_action = np.tile(wrapped_action_space.low, horizon)
high_action = np.tile(wrapped_action_space.high, horizon)
low = np.concatenate((low_obs, low_action))
high = np.concatenate((high_obs, high_action))
# Overwrite the observation space
env.observation_space = spaces.Box(low=low, high=high, dtype=wrapped_obs_space.dtype) # type: ignore[arg-type]
super().__init__(env)
self.horizon = horizon
self.low_action, self.high_action = low_action, high_action
self.low_obs, self.high_obs = low_obs, high_obs
self.low, self.high = low, high
self.obs_history = np.zeros(low_obs.shape, low_obs.dtype)
self.action_history = np.zeros(low_action.shape, low_action.dtype)
def _create_obs_from_history(self) -> np.ndarray:
return np.concatenate((self.obs_history, self.action_history))
def reset(self, seed: int | None = None, options: dict | None = None) -> tuple[np.ndarray, dict]:
# Flush the history
self.obs_history[...] = 0
self.action_history[...] = 0
assert options is None, "Options not supported for now"
obs, info = self.env.reset(seed=seed)
self.obs_history[..., -obs.shape[-1] :] = obs
return self._create_obs_from_history(), info
def step(self, action: np.ndarray) -> tuple[np.ndarray, SupportsFloat, bool, bool, dict]:
obs, reward, terminated, truncated, info = self.env.step(action)
last_ax_size = obs.shape[-1]
self.obs_history = np.roll(self.obs_history, shift=-last_ax_size, axis=-1)
self.obs_history[..., -obs.shape[-1] :] = obs
self.action_history = np.roll(self.action_history, shift=-action.shape[-1], axis=-1)
self.action_history[..., -action.shape[-1] :] = action
return self._create_obs_from_history(), reward, terminated, truncated, info
class HistoryWrapperObsDict(gym.Wrapper):
"""History Wrapper for dict observation spaces.
:param env: Gym environment with a dict observation space.
:param horizon: Number of past steps to include.
"""
def __init__(self, env: gym.Env, horizon: int = 2) -> None:
assert isinstance(env.observation_space, spaces.Dict)
assert isinstance(env.observation_space.spaces["observation"], spaces.Box)
assert isinstance(env.action_space, spaces.Box)
wrapped_obs_space = env.observation_space.spaces["observation"]
wrapped_action_space = env.action_space
low_obs = np.tile(wrapped_obs_space.low, horizon)
high_obs = np.tile(wrapped_obs_space.high, horizon)
low_action = np.tile(wrapped_action_space.low, horizon)
high_action = np.tile(wrapped_action_space.high, horizon)
low = np.concatenate((low_obs, low_action))
high = np.concatenate((high_obs, high_action))
# Overwrite the observation space
env.observation_space.spaces["observation"] = spaces.Box(
low=low,
high=high,
dtype=wrapped_obs_space.dtype, # type: ignore[arg-type]
)
super().__init__(env)
self.horizon = horizon
self.low_action, self.high_action = low_action, high_action
self.low_obs, self.high_obs = low_obs, high_obs
self.low, self.high = low, high
self.obs_history = np.zeros(low_obs.shape, low_obs.dtype)
self.action_history = np.zeros(low_action.shape, low_action.dtype)
def _create_obs_from_history(self) -> np.ndarray:
return np.concatenate((self.obs_history, self.action_history))
def reset(self, seed: int | None = None, options: dict | None = None) -> tuple[dict[str, np.ndarray], dict]:
# Flush the history
self.obs_history[...] = 0
self.action_history[...] = 0
assert options is None, "Options not supported for now"
obs_dict, info = self.env.reset(seed=seed)
obs = obs_dict["observation"]
self.obs_history[..., -obs.shape[-1] :] = obs
obs_dict["observation"] = self._create_obs_from_history()
return obs_dict, info
def step(self, action: np.ndarray) -> tuple[dict[str, np.ndarray], SupportsFloat, bool, bool, dict]:
obs_dict, reward, terminated, truncated, info = self.env.step(action)
obs = obs_dict["observation"]
last_ax_size = obs.shape[-1]
self.obs_history = np.roll(self.obs_history, shift=-last_ax_size, axis=-1)
self.obs_history[..., -obs.shape[-1] :] = obs
self.action_history = np.roll(self.action_history, shift=-action.shape[-1], axis=-1)
self.action_history[..., -action.shape[-1] :] = action
obs_dict["observation"] = self._create_obs_from_history()
return obs_dict, reward, terminated, truncated, info
|