File size: 12,316 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
from typing import Any, Union, List
from collections import namedtuple
from easydict import EasyDict
import gym
import copy
import numpy as np

from overcooked_ai_py.mdp.actions import Action, Direction
from overcooked_ai_py.mdp.overcooked_mdp import PlayerState, OvercookedGridworld, OvercookedState, ObjectState, \
    SoupState, Recipe
from overcooked_ai_py.mdp.overcooked_env import OvercookedEnv, DEFAULT_ENV_PARAMS

from ding.envs import BaseEnv
from ding.utils import ENV_REGISTRY, deep_merge_dicts

OvercookEnvTimestep = namedtuple('OvercookEnvTimestep', ['obs', 'reward', 'done', 'info'])

# n, s = Direction.NORTH, Direction.SOUTH
# e, w = Direction.EAST, Direction.WEST
# stay, interact = Action.STAY, Action.INTERACT
# Action.ALL_ACTIONS: [n, s, e, w, stay, interact]


@ENV_REGISTRY.register('overcooked')
class OvercookEnv(BaseEnv):
    config = EasyDict(
        dict(
            env_name="cramped_room",
            horizon=400,
            concat_obs=False,
            action_mask=True,
            shape_reward=True,
        )
    )

    def __init__(self, cfg) -> None:
        self._cfg = deep_merge_dicts(self.config, cfg)
        self._env_name = self._cfg.env_name
        self._horizon = self._cfg.horizon
        self._concat_obs = self._cfg.concat_obs
        self._action_mask = self._cfg.action_mask
        self._shape_reward = self._cfg.shape_reward
        self.mdp = OvercookedGridworld.from_layout_name(self._env_name)
        self.base_env = OvercookedEnv.from_mdp(self.mdp, horizon=self._horizon, info_level=0)

        # rightnow overcook environment encoding only support 2 agent game
        self.agent_num = 2
        self.action_dim = len(Action.ALL_ACTIONS)
        self.action_space = gym.spaces.Discrete(len(Action.ALL_ACTIONS))
        # set up obs shape
        featurize_fn = lambda mdp, state: mdp.lossless_state_encoding(state)
        self.featurize_fn = featurize_fn
        dummy_mdp = self.base_env.mdp
        dummy_state = dummy_mdp.get_standard_start_state()
        obs_shape = self.featurize_fn(dummy_mdp, dummy_state)[0].shape  # (5, 4, 26)
        obs_shape = (obs_shape[-1], *obs_shape[:-1])  # permute channel first
        if self._concat_obs:
            obs_shape = (obs_shape[0] * 2, *obs_shape[1:])
        else:
            obs_shape = (2, ) + obs_shape
        self.observation_space = gym.spaces.Box(low=0, high=1, shape=obs_shape, dtype=np.int64)
        if self._action_mask:
            self.observation_space = gym.spaces.Dict(
                {
                    'agent_state': self.observation_space,
                    'action_mask': gym.spaces.Box(
                        low=0, high=1, shape=(self.agent_num, self.action_dim), dtype=np.int64
                    )
                }
            )
        self.reward_space = gym.spaces.Box(low=0, high=100, shape=(1, ), dtype=np.float32)

    def seed(self, seed: int, dynamic_seed: bool = True) -> None:
        self._seed = seed
        self._dynamic_seed = dynamic_seed
        np.random.seed(self._seed)

    def close(self) -> None:
        # Note: the real env instance only has a empty close method, only pas
        pass

    def random_action(self):
        return [self.action_space.sample() for _ in range(self.agent_num)]

    def step(self, action):
        assert all(self.action_space.contains(a) for a in action), "%r (%s) invalid" % (action, type(action))
        agent_action, other_agent_action = [Action.INDEX_TO_ACTION[a] for a in action]

        if self.agent_idx == 0:
            joint_action = (agent_action, other_agent_action)
        else:
            joint_action = (other_agent_action, agent_action)

        next_state, reward, done, env_info = self.base_env.step(joint_action)
        reward = np.array([float(reward)])
        self._eval_episode_return += reward
        if self._shape_reward:
            self._eval_episode_return += sum(env_info['shaped_r_by_agent'])
            reward += sum(env_info['shaped_r_by_agent'])

        ob_p0, ob_p1 = self.featurize_fn(self.mdp, next_state)
        ob_p0, ob_p1 = self.obs_preprocess(ob_p0), self.obs_preprocess(ob_p1)
        if self.agent_idx == 0:
            both_agents_ob = [ob_p0, ob_p1]
        else:
            both_agents_ob = [ob_p1, ob_p0]
        if self._concat_obs:
            both_agents_ob = np.concatenate(both_agents_ob)
        else:
            both_agents_ob = np.stack(both_agents_ob)

        env_info["policy_agent_idx"] = self.agent_idx
        env_info["eval_episode_return"] = self._eval_episode_return
        env_info["other_agent_env_idx"] = 1 - self.agent_idx

        action_mask = self.get_action_mask()
        if self._action_mask:
            obs = {
                "agent_state": both_agents_ob,
                # "overcooked_state": self.base_env.state,
                "action_mask": action_mask
            }
        else:
            obs = both_agents_ob
        return OvercookEnvTimestep(obs, reward, done, env_info)

    def obs_preprocess(self, obs):
        obs = obs.transpose(2, 0, 1)
        return obs

    def reset(self):
        self.base_env.reset()
        self._eval_episode_return = 0
        self.mdp = self.base_env.mdp
        # random init agent index
        self.agent_idx = np.random.choice([0, 1])
        ob_p0, ob_p1 = self.featurize_fn(self.mdp, self.base_env.state)
        ob_p0, ob_p1 = self.obs_preprocess(ob_p0), self.obs_preprocess(ob_p1)

        if self.agent_idx == 0:
            both_agents_ob = [ob_p0, ob_p1]
        else:
            both_agents_ob = [ob_p1, ob_p0]
        if self._concat_obs:
            both_agents_ob = np.concatenate(both_agents_ob)
        else:
            both_agents_ob = np.stack(both_agents_ob)

        action_mask = self.get_action_mask()

        if self._action_mask:
            obs = {"agent_state": both_agents_ob, "action_mask": action_mask}
        else:
            obs = both_agents_ob
        return obs

    def get_available_actions(self):
        return self.mdp.get_actions(self.base_env.state)

    def get_action_mask(self):
        available_actions = self.get_available_actions()

        action_masks = np.zeros((self.agent_num, self.action_dim)).astype(np.int64)

        for i in range(self.action_dim):
            if Action.INDEX_TO_ACTION[i] in available_actions[0]:
                action_masks[0][i] = 1
            if Action.INDEX_TO_ACTION[i] in available_actions[1]:
                action_masks[1][i] = 1

        return action_masks

    def __repr__(self):
        return "DI-engine Overcooked Env"


@ENV_REGISTRY.register('overcooked_game')
class OvercookGameEnv(BaseEnv):
    config = EasyDict(
        dict(
            env_name="cramped_room",
            horizon=400,
            concat_obs=False,
            action_mask=False,
            shape_reward=True,
        )
    )

    def __init__(self, cfg) -> None:
        self._cfg = deep_merge_dicts(self.config, cfg)
        self._env_name = self._cfg.env_name
        self._horizon = self._cfg.horizon
        self._concat_obs = self._cfg.concat_obs
        self._action_mask = self._cfg.action_mask
        self._shape_reward = self._cfg.shape_reward
        self.mdp = OvercookedGridworld.from_layout_name(self._env_name)
        self.base_env = OvercookedEnv.from_mdp(self.mdp, horizon=self._horizon, info_level=0)

        # rightnow overcook environment encoding only support 2 agent game
        self.agent_num = 2
        self.action_dim = len(Action.ALL_ACTIONS)
        self.action_space = gym.spaces.Discrete(len(Action.ALL_ACTIONS))
        # set up obs shape
        featurize_fn = lambda mdp, state: mdp.lossless_state_encoding(state)
        self.featurize_fn = featurize_fn
        dummy_mdp = self.base_env.mdp
        dummy_state = dummy_mdp.get_standard_start_state()
        obs_shape = self.featurize_fn(dummy_mdp, dummy_state)[0].shape  # (5, 4, 26)
        obs_shape = (obs_shape[-1], *obs_shape[:-1])  # permute channel first
        if self._concat_obs:
            obs_shape = (obs_shape[0] * 2, *obs_shape[1:])
        else:
            obs_shape = (2, ) + obs_shape
        self.observation_space = gym.spaces.Box(low=0, high=1, shape=obs_shape, dtype=np.int64)
        if self._action_mask:
            self.observation_space = gym.spaces.Dict(
                {
                    'agent_state': self.observation_space,
                    'action_mask': gym.spaces.Box(
                        low=0, high=1, shape=(self.agent_num, self.action_dim), dtype=np.int64
                    )
                }
            )

        self.reward_space = gym.spaces.Box(low=0, high=100, shape=(1, ), dtype=np.float32)

    def seed(self, seed: int, dynamic_seed: bool = True) -> None:
        self._seed = seed
        self._dynamic_seed = dynamic_seed
        np.random.seed(self._seed)

    def close(self) -> None:
        # Note: the real env instance only has a empty close method, only pass
        pass

    def random_action(self):
        return [self.action_space.sample() for _ in range(self.agent_num)]

    def step(self, action):
        assert all(self.action_space.contains(a) for a in action), "%r (%s) invalid" % (action, type(action))
        agent_action, other_agent_action = [Action.INDEX_TO_ACTION[a] for a in action]

        if self.agent_idx == 0:
            joint_action = (agent_action, other_agent_action)
        else:
            joint_action = (other_agent_action, agent_action)

        next_state, reward, done, env_info = self.base_env.step(joint_action)

        reward = np.array([float(reward)])
        self._eval_episode_return += reward
        if self._shape_reward:
            self._eval_episode_return += sum(env_info['shaped_r_by_agent'])
            reward += sum(env_info['shaped_r_by_agent'])
        ob_p0, ob_p1 = self.featurize_fn(self.mdp, next_state)
        ob_p0, ob_p1 = self.obs_preprocess(ob_p0), self.obs_preprocess(ob_p1)
        if self.agent_idx == 0:
            both_agents_ob = [ob_p0, ob_p1]
        else:
            both_agents_ob = [ob_p1, ob_p0]
        if self._concat_obs:
            both_agents_ob = np.concatenate(both_agents_ob)
        else:
            both_agents_ob = np.stack(both_agents_ob)

        env_info["policy_agent_idx"] = self.agent_idx
        env_info["eval_episode_return"] = self._eval_episode_return
        env_info["other_agent_env_idx"] = 1 - self.agent_idx

        action_mask = self.get_action_mask()
        if self._action_mask:
            obs = {"agent_state": both_agents_ob, "action_mask": action_mask}
        else:
            obs = both_agents_ob
        return OvercookEnvTimestep(obs, reward, done, env_info)

    def obs_preprocess(self, obs):
        obs = obs.transpose(2, 0, 1)
        return obs

    def reset(self):
        self.base_env.reset()
        self._eval_episode_return = 0
        self.mdp = self.base_env.mdp
        # random init agent index
        self.agent_idx = np.random.choice([0, 1])
        #fix init agent index
        self.agent_idx = 0
        ob_p0, ob_p1 = self.featurize_fn(self.mdp, self.base_env.state)
        ob_p0, ob_p1 = self.obs_preprocess(ob_p0), self.obs_preprocess(ob_p1)

        if self.agent_idx == 0:
            both_agents_ob = [ob_p0, ob_p1]
        else:
            both_agents_ob = [ob_p1, ob_p0]
        if self._concat_obs:
            both_agents_ob = np.concatenate(both_agents_ob)
        else:
            both_agents_ob = np.stack(both_agents_ob)

        action_mask = self.get_action_mask()

        if self._action_mask:
            obs = {"agent_state": both_agents_ob, "action_mask": action_mask}
        else:
            obs = both_agents_ob
        return obs

    def get_available_actions(self):
        return self.mdp.get_actions(self.base_env.state)

    def get_action_mask(self):
        available_actions = self.get_available_actions()

        action_masks = np.zeros((self.agent_num, self.action_dim)).astype(np.int64)

        for i in range(self.action_dim):
            if Action.INDEX_TO_ACTION[i] in available_actions[0]:
                action_masks[0][i] = 1
            if Action.INDEX_TO_ACTION[i] in available_actions[1]:
                action_masks[1][i] = 1

        return action_masks

    def __repr__(self):
        return "DI-engine Overcooked GameEnv"