File size: 13,004 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
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
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
from gym_minigrid.minigrid import *
from gym_minigrid.register import register
 

class Wizard(NPC):
    """
    A simple NPC that knows who is telling the truth
    """

    def __init__(self, color, name, env):
        super().__init__(color)
        self.name = name
        self.env = env
        self.npc_type = 0  # this will be put into the encoding

    def listen(self, utterance):
        if utterance == TalkHardSesameNPCGuidesGrammar.construct_utterance([0, 1]):
            return "Ask {}.".format(self.env.true_guide.name)

        return None

    def is_near_agent(self):
        ax, ay = self.env.agent_pos
        wx, wy = self.cur_pos
        if (ax == wx and abs(ay - wy) == 1) or (ay == wy and abs(ax - wx) == 1):
            return True
        return False


class Guide(NPC):
    """
    A simple NPC that knows the correct door.
    """

    def __init__(self, color, name, env, liar=False):
        super().__init__(color)
        self.name = name
        self.env = env
        self.liar = liar
        self.npc_type = 1  # this will be put into the encoding

        # Select a random target object as mission
        obj_idx = self.env._rand_int(0, len(self.env.door_pos))
        self.target_pos = self.env.door_pos[obj_idx]
        self.target_color = self.env.door_colors[obj_idx]

    def listen(self, utterance):
        if utterance == TalkHardSesameNPCGuidesGrammar.construct_utterance([0, 1]):
            if self.liar:
                fake_colors = [c for c in self.env.door_colors if c != self.env.target_color]
                fake_color = self.env._rand_elem(fake_colors)

                # Generate the mission string
                assert fake_color != self.env.target_color
                return 'go to the %s door' % fake_color

            else:
                return self.env.mission

        return None

    def render(self, img):
        c = COLORS[self.color]

        # Draw eyes
        fill_coords(img, point_in_circle(cx=0.70, cy=0.50, r=0.10), c)
        fill_coords(img, point_in_circle(cx=0.30, cy=0.50, r=0.10), c)

        # Draw mouth
        fill_coords(img, point_in_rect(0.20, 0.80, 0.72, 0.81), c)

        # #Draw hat
        # tri_fn = point_in_triangle(
        #     (0.15, 0.25),
        #     (0.85, 0.25),
        #     (0.50, 0.05),
        # )
        # fill_coords(img, tri_fn, c)

    def is_near_agent(self):
        ax, ay = self.env.agent_pos
        wx, wy = self.cur_pos
        if (ax == wx and abs(ay - wy) == 1) or (ay == wy and abs(ax - wx) == 1):
            return True
        return False


class TalkHardSesameNPCGuidesGrammar(object):

    templates = ["Where is", "Open"]
    things = ["sesame", "the exit"]

    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 GoToDoorTalkHardSesameNPCGuidesEnv(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,
        hear_yourself=False,
        diminished_reward=True,
        step_penalty=False
    ):
        assert size >= 5

        super().__init__(
            grid_size=size,
            max_steps=5*size**2,
            # Set this to True for maximum speed
            see_through_walls=True,
            actions=MiniGridEnv.Actions,
            action_space=spaces.MultiDiscrete([
                len(MiniGridEnv.Actions),
                *TalkHardSesameNPCGuidesGrammar.grammar_action_space.nvec
            ])
        )
        self.hear_yourself = hear_yourself
        self.diminished_reward = diminished_reward
        self.step_penalty = step_penalty

        self.empty_symbol = "NA \n"

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

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

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

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

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

        # Generate the 4 doors at random positions
        self.door_pos = []
        self.door_front_pos = []  # Remembers positions in front of door to avoid setting wizard here

        self.door_pos.append((self._rand_int(2, width-2), 0))
        self.door_front_pos.append((self.door_pos[-1][0], self.door_pos[-1][1]+1))

        self.door_pos.append((self._rand_int(2, width-2), height-1))
        self.door_front_pos.append((self.door_pos[-1][0], self.door_pos[-1][1] - 1))

        self.door_pos.append((0, self._rand_int(2, height-2)))
        self.door_front_pos.append((self.door_pos[-1][0] + 1, self.door_pos[-1][1]))

        self.door_pos.append((width-1, self._rand_int(2, height-2)))
        self.door_front_pos.append((self.door_pos[-1][0] - 1, self.door_pos[-1][1]))

        # Generate the door colors
        self.door_colors = []
        while len(self.door_colors) < len(self.door_pos):
            color = self._rand_elem(COLOR_NAMES)
            if color in self.door_colors:
                continue
            self.door_colors.append(color)

        # Place the doors in the grid
        for idx, pos in enumerate(self.door_pos):
            color = self.door_colors[idx]
            self.grid.set(*pos, Door(color))


        # Set a randomly coloured WIZARD at a random position
        color = self._rand_elem(COLOR_NAMES)
        self.wizard = Wizard(color, "Gandalf", self)

        # Place it randomly, omitting front of door positions
        self.place_obj(self.wizard,
                       size=(width, height),
                       reject_fn=lambda _, p: tuple(p) in self.door_front_pos)


        # add guides
        GUIDE_NAMES = ["John", "Jack"]

        # Set a randomly coloured TRUE GUIDE at a random position
        name = self._rand_elem(GUIDE_NAMES)
        color = self._rand_elem(COLOR_NAMES)
        self.true_guide = Guide(color, name, self, liar=False)

        # Place it randomly, omitting invalid positions
        self.place_obj(self.true_guide,
                       size=(width, height),
                       # reject_fn=lambda _, p: tuple(p) in self.door_front_pos)
                       reject_fn=lambda _, p: tuple(p) in [*self.door_front_pos, tuple(self.wizard.cur_pos)])

        # Set a randomly coloured FALSE GUIDE at a random position
        name = self._rand_elem([n for n in GUIDE_NAMES if n != self.true_guide.name])
        color = self._rand_elem(COLOR_NAMES)
        self.false_guide = Guide(color, name, self, liar=True)

        # Place it randomly, omitting invalid positions
        self.place_obj(self.false_guide,
                       size=(width, height),
                       reject_fn=lambda _, p: tuple(p) in [
                           *self.door_front_pos, tuple(self.wizard.cur_pos), tuple(self.true_guide.cur_pos)])
        assert self.true_guide.name != self.false_guide.name

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

        # Select a random target door
        self.doorIdx = self._rand_int(0, len(self.door_pos))
        self.target_pos = self.door_pos[self.doorIdx]
        self.target_color = self.door_colors[self.doorIdx]

        # Generate the mission string
        self.mission = 'go to the %s door' % self.target_color

        # 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 = ""
        
        self.conversation = self.utterance

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

        # assert all nan or neither nan
        assert len(set(np.isnan(utterance_action))) == 1

        speak_flag = not all(np.isnan(utterance_action))

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

        if speak_flag:
            utterance = TalkHardSesameNPCGuidesGrammar.construct_utterance(utterance_action)
            if self.hear_yourself:
                self.utterance += "YOU: {} \n".format(utterance)
                
            self.conversation += "YOU: {} \n".format(utterance)

            # check if near wizard
            if hasattr(self, "wizard"):
                if self.wizard.is_near_agent():
                    reply = self.wizard.listen(utterance)

                    if reply:
                        self.utterance += "{}: {} \n".format(self.wizard.name, reply)
                        self.conversation += "{}: {} \n".format(self.wizard.name, reply)

            if self.true_guide.is_near_agent():
                reply = self.true_guide.listen(utterance)

                if reply:
                    self.utterance += "{}: {} \n".format(self.true_guide.name, reply)
                    self.conversation += "{}: {} \n".format(self.true_guide.name, reply)

            if hasattr(self, "false_guide"):
                if self.false_guide.is_near_agent():
                    reply = self.false_guide.listen(utterance)

                    if reply:
                        self.utterance += "{}: {} \n".format(self.false_guide.name, reply)
                        self.conversation += "{}: {} \n".format(self.false_guide.name, reply)

            if utterance == TalkHardSesameNPCGuidesGrammar.construct_utterance([1, 0]):
                ax, ay = self.agent_pos
                tx, ty = self.target_pos

                if (ax == tx and abs(ay - ty) == 1) or (ay == ty and abs(ax - tx) == 1):
                    reward = self._reward()

                for dx, dy in self.door_pos:
                    if (ax == dx and abs(ay - dy) == 1) or (ay == dy and abs(ax - dx) == 1):
                        # agent has chosen some door episode, regardless of if the door is correct the episode is over
                        done = True

        # Don't let the agent open any of the doors
        if p_action == self.actions.toggle:
            done = True

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

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

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

        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)
        print(self.conversation)
        self.window.set_caption(self.conversation, [
            "Gandalf:",
            "Jack:",
            "John:",
            "Where is the exit",
            "Open sesame",
        ])
        return obs



class GoToDoorTalkHardSesameNPCGuides8x8Env(GoToDoorTalkHardSesameNPCGuidesEnv):
    def __init__(self):
        super().__init__(size=8)


class GoToDoorTalkHardSesameNPCGuides6x6Env(GoToDoorTalkHardSesameNPCGuidesEnv):
    def __init__(self):
        super().__init__(size=6)


# hear yourself
class GoToDoorTalkHardSesameNPCGuidesHY8x8Env(GoToDoorTalkHardSesameNPCGuidesEnv):
    def __init__(self):
        super().__init__(size=8, hear_yourself=True)


class GoToDoorTalkHardSesameNPCGuidesHY6x6Env(GoToDoorTalkHardSesameNPCGuidesEnv):
    def __init__(self):
        super().__init__(size=6, hear_yourself=True)


class GoToDoorTalkHardSesameNPCGuidesHY5x5Env(GoToDoorTalkHardSesameNPCGuidesEnv):
    def __init__(self):
        super().__init__(size=5, hear_yourself=True)

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

register(
    id='MiniGrid-GoToDoorTalkHardSesameNPCGuides-6x6-v0',
    entry_point='gym_minigrid.envs:GoToDoorTalkHardSesameNPCGuides6x6Env'
)

register(
    id='MiniGrid-GoToDoorTalkHardSesameNPCGuides-8x8-v0',
    entry_point='gym_minigrid.envs:GoToDoorTalkHardSesameNPCGuides8x8Env'
)
register(
    id='MiniGrid-GoToDoorTalkHardSesameNPCGuidesHY-5x5-v0',
    entry_point='gym_minigrid.envs:GoToDoorTalkHardSesameNPCGuidesHY5x5Env'
)

register(
    id='MiniGrid-GoToDoorTalkHardSesameNPCGuidesHY-6x6-v0',
    entry_point='gym_minigrid.envs:GoToDoorTalkHardSesameNPGuidesCHY6x6Env'
)

register(
    id='MiniGrid-GoToDoorTalkHardSesameNPCGuidesHY-8x8-v0',
    entry_point='gym_minigrid.envs:GoToDoorTalkHardSesameNPCGuidesHY8x8Env'
)