File size: 9,350 Bytes
be5548b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np

from gym_minigrid.minigrid import *
from gym_minigrid.register import register
 
import time
from collections import deque


class Peer(NPC):
    """
    A dancing NPC that the agent has to copy
    """

    def __init__(self, color, name, env):
        super().__init__(color)
        self.name = name
        self.npc_dir = 1  # NPC initially looks downward
        self.npc_type = 0
        self.env = env
        self.npc_actions = []
        self.dancing_step_idx = 0
        self.actions = MiniGridEnv.Actions
        self.add_npc_direction = True
        self.available_moves = [self.rotate_left, self.rotate_right, self.go_forward, self.toggle_action]

        selected_door_id = self.env._rand_elem([0, 1])
        self.selected_door_pos = [self.env.door_pos_top, self.env.door_pos_bottom][selected_door_id]
        self.selected_door = [self.env.door_top, self.env.door_bottom][selected_door_id]
        self.joint_attention_achieved = False

    def can_overlap(self):
        # If the NPC is hidden, agent can overlap on it
        return self.env.hidden_npc

    def encode(self, nb_dims=3):
        if self.env.hidden_npc:
            if nb_dims == 3:
                return (1, 0, 0)
            elif nb_dims == 4:
                return (1, 0, 0, 0)
        else:
            return super().encode(nb_dims=nb_dims)

    def step(self):

        distance_to_door = np.abs(self.selected_door_pos - self.cur_pos).sum(-1)

        if all(self.front_pos == self.selected_door_pos) and self.selected_door.is_open:
            # in front of door
            self.go_forward()

        elif distance_to_door == 1 and not self.joint_attention_achieved:
            # before turning to the door look at the agent
            wanted_dir = self.compute_wanted_dir(self.env.agent_pos)
            act = self.compute_turn_action(wanted_dir)
            act()
            if self.is_eye_contact():
                self.joint_attention_achieved = True

        else:
            act = self.path_to_toggle_pos(self.selected_door_pos)
            act()

        # not really important as the NPC doesn't speak
        if self.env.hidden_npc:
            return None



class HelperGrammar(object):

    templates = ["Move your", "Shake your"]
    things = ["body", "head"]

    grammar_action_space = spaces.MultiDiscrete([len(templates), len(things)])

    @classmethod
    def construct_utterance(cls, action):
        return cls.templates[int(action[0])] + " " + cls.things[int(action[1])] + " "


class HelperEnv(MultiModalMiniGridEnv):
    """
    Environment in which the agent is instructed to go to a given object
    named using an English text string
    """

    def __init__(
        self,
        size=5,
        diminished_reward=True,
        step_penalty=False,
        knowledgeable=False,
        max_steps=20,
        hidden_npc=False,
    ):
        assert size >= 5
        self.empty_symbol = "NA \n"
        self.diminished_reward = diminished_reward
        self.step_penalty = step_penalty
        self.knowledgeable = knowledgeable
        self.hidden_npc = hidden_npc

        super().__init__(
            grid_size=size,
            max_steps=max_steps,
            # Set this to True for maximum speed
            see_through_walls=True,
            actions=MiniGridEnv.Actions,
            action_space=spaces.MultiDiscrete([
                len(MiniGridEnv.Actions),
                *HelperGrammar.grammar_action_space.nvec
            ]),
            add_npc_direction=True
        )

        print({
            "size": size,
            "diminished_reward": diminished_reward,
            "step_penalty": step_penalty,
        })

    def _gen_grid(self, width, height):
        # Create the grid
        self.grid = Grid(width, height, nb_obj_dims=4)

        # Randomly vary the room width and height
        width = self._rand_int(5, width+1)
        height = self._rand_int(5, height+1)

        self.wall_x = width-1
        self.wall_y = height-1

        # Generate the surrounding walls
        self.grid.wall_rect(0, 0, width, height)

        # add lava
        self.grid.vert_wall(width//2, 1, height - 2, Lava)

        # door top
        door_color_top = self._rand_elem(COLOR_NAMES)
        self.door_pos_top = (width-1, 1)
        self.door_top = Door(door_color_top, is_locked=True)
        self.grid.set(*self.door_pos_top, self.door_top)

        # switch top
        self.switch_pos_top = (0, 1)
        self.switch_top = Switch(door_color_top, lockable_object=self.door_top, locker_switch=True)
        self.grid.set(*self.switch_pos_top, self.switch_top)

        # door bottom
        door_color_bottom = self._rand_elem(COLOR_NAMES)
        self.door_pos_bottom = (width-1, height-2)
        self.door_bottom = Door(door_color_bottom, is_locked=True)
        self.grid.set(*self.door_pos_bottom, self.door_bottom)

        # switch bottom
        self.switch_pos_bottom = (0, height-2)
        self.switch_bottom = Switch(door_color_bottom, lockable_object=self.door_bottom, locker_switch=True)
        self.grid.set(*self.switch_pos_bottom, self.switch_bottom)

        # save to variables
        self.switches = [self.switch_top, self.switch_bottom]
        self.switches_pos = [self.switch_pos_top, self.switch_pos_bottom]
        self.door = [self.door_top, self.door_bottom]
        self.door_pos = [self.door_pos_top, self.door_pos_bottom]

        # Set a randomly coloured Dancer NPC
        color = self._rand_elem(COLOR_NAMES)
        self.peer = Peer(color, "Jill", self)

        # Place it on the middle right side of the room
        peer_pos = np.array((self._rand_int(width//2+1, width - 1), self._rand_int(1, height - 1)))

        self.grid.set(*peer_pos, self.peer)
        self.peer.init_pos = peer_pos
        self.peer.cur_pos = peer_pos

        # Randomize the agent's start position and orientation
        self.place_agent(size=(width//2, height))

        # Generate the mission string
        self.mission = 'watch dancer and repeat his moves afterwards'

        # Dummy beginning string
        self.beginning_string = "This is what you hear. \n"
        self.utterance = self.beginning_string

        # utterance appended at the end of each step
        self.utterance_history = ""

        # used for rendering
        self.conversation = self.utterance
        self.outcome_info = None

    def step(self, action):
        p_action = action[0]
        utterance_action = action[1:]

        obs, reward, done, info = super().step(p_action)
        self.peer.step()

        if np.isnan(p_action):
            pass

        if p_action == self.actions.done:
            done = True

        elif all(self.agent_pos == self.door_pos_top):
            done = True

        elif all(self.agent_pos == self.door_pos_bottom):
            done = True

        elif all([self.switch_top.is_on, self.switch_bottom.is_on]):
            # if both switches are on no reward is given and episode ends
            done = True

        elif all(self.peer.cur_pos == self.peer.selected_door_pos):
            reward = self._reward()
            done = True

        # discount
        if self.step_penalty:
            reward = reward - 0.01

        if self.hidden_npc:
            # all npc are hidden
            assert np.argwhere(obs['image'][:,:,0] == OBJECT_TO_IDX['npc']).size == 0
            assert "{}:".format(self.peer.name) not in self.utterance

        # fill observation with text
        self.append_existing_utterance_to_history()
        obs = self.add_utterance_to_observation(obs)
        self.reset_utterance()

        if done:
            if reward > 0:
                self.outcome_info = "SUCCESS: agent got {} reward \n".format(np.round(reward, 1))
            else:
                self.outcome_info = "FAILURE: agent got {} reward \n".format(reward)

        return obs, reward, done, info

    def _reward(self):
        if self.diminished_reward:
            return super()._reward()
        else:
            return 1.0

    def render(self, *args, **kwargs):
        obs = super().render(*args, **kwargs)
        self.window.clear_text()  # erase previous text

        # self.window.set_caption(self.conversation, [self.peer.name])
        # self.window.ax.set_title("correct door: {}".format(self.true_guide.target_color), loc="left", fontsize=10)
        if self.outcome_info:
            color = None
            if "SUCCESS" in self.outcome_info:
                color = "lime"
            elif "FAILURE" in self.outcome_info:
                color = "red"
            self.window.add_text(*(0.01, 0.85, self.outcome_info),
                                 **{'fontsize':15, 'color':color, 'weight':"bold"})

        self.window.show_img(obs)  # re-draw image to add changes to window
        return obs


class Helper8x8Env(HelperEnv):
    def __init__(self, **kwargs):
        super().__init__(size=8, max_steps=20, **kwargs)


class Helper6x6Env(HelperEnv):
    def __init__(self):
        super().__init__(size=6, max_steps=20)



register(
    id='MiniGrid-Helper-5x5-v0',
    entry_point='gym_minigrid.envs:HelperEnv'
)

register(
    id='MiniGrid-Helper-6x6-v0',
    entry_point='gym_minigrid.envs:Helper6x6Env'
)

register(
    id='MiniGrid-Helper-8x8-v0',
    entry_point='gym_minigrid.envs:Helper8x8Env'
)