/
karel.py
496 lines (437 loc) · 18.6 KB
/
karel.py
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
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
import os
import numpy as np
import scipy
from scipy import spatial
MAX_NUM_MARKER = 10
state_table = {
0: 'Karel facing North',
1: 'Karel facing East',
2: 'Karel facing South',
3: 'Karel facing West',
4: 'Wall',
5: '0 marker',
6: '1 marker',
7: '2 markers',
8: '3 markers',
9: '4 markers',
10: '5 markers',
11: '6 markers',
12: '7 markers',
13: '8 markers',
14: '9 markers',
15: '10 markers'
}
action_table = {
0: 'Move',
1: 'Turn left',
2: 'Turn right',
3: 'Pick up a marker',
4: 'Put a marker'
}
class Karel_world(object):
def __init__(self, s=None, make_error=True, env_task="program", task_definition='program' ,reward_diff=False, final_reward_scale=True):
if s is not None:
self.set_new_state(s)
self.make_error = make_error
self.env_task = env_task
self.task_definition = task_definition
self.rescale_reward = True
self.final_reward_scale = final_reward_scale
self.reward_diff = reward_diff
self.num_actions = len(action_table)
def set_new_state(self, s, metadata=None):
self.perception_count = 0
self.s = s.astype(np.bool)
self.s_h = [self.s.copy()]
self.a_h = []
self.h = self.s.shape[0]
self.w = self.s.shape[1]
p_v = self.get_perception_vector()
self.p_v_h = [p_v.copy()]
if self.task_definition != "program":
self.r_h = []
self.d_h = []
self.program_reward = 0.0
self.prev_pos_reward = 0.0
self.done = False
self.metadata = metadata
self.total_markers = np.sum(s[:,:,6:])
###################################
### Collect Demonstrations ###
###################################
def clear_history(self):
self.perception_count = 0
self.s_h = [self.s.copy()]
self.a_h = []
self.p_v_h = []
if self.task_definition != "program":
self.r_h = []
self.d_h = []
self.program_reward = 0.0
self.prev_pos_reward = 0.0
self.done = False
self.total_markers = np.sum(self.s_h[-1][:,:,6:])
def add_to_history(self, a_idx, agent_pos, made_error=False):
self.s_h.append(self.s.copy())
self.a_h.append(a_idx)
p_v = self.get_perception_vector()
self.p_v_h.append(p_v.copy())
if self.task_definition != "program":
reward, done = self._get_state_reward(agent_pos, made_error)
self.done = self.done or done
self.r_h.append(reward)
self.d_h.append(done)
self.program_reward += reward
if self.task_definition != 'program' and not made_error:
if a_idx == 3: self.total_markers -= 1
if a_idx == 4: self.total_markers += 1
def _get_cleanHouse_task_reward(self, agent_pos):
done = False
reward = 0
w = self.w
h = self.h
state = self.s_h[-1]
for mpos in self.metadata['marker_positions']:
if state[mpos[0], mpos[1], 5] and not state[mpos[0], mpos[1], 6]:
reward += 1
reward = reward / len(self.metadata['marker_positions'])
done = reward == 1
reward = reward if self.env_task == 'cleanHouse' else float(done)
self.done = self.done or done
return reward, done
def _get_harvester_task_reward(self, agent_pos):
done = False
reward = 0
w = self.w
h = self.h
state = self.s_h[-1]
max_markers = (w-2)*(h-2)
reward = ( max_markers - self.total_markers ) / max_markers
done = reward == 1
reward = reward if self.env_task == 'harvester' else float(done)
self.done = self.done or done
return reward, done
def _get_randomMaze_task_reward(self, agent_pos):
done = False
reward = 0
w = self.w
h = self.h
state = self.s_h[-1]
# initial marker position
init_state = self.s_h[0]
x, y = np.where(init_state[:, :, 6] > 0)
if len(x) != 1: assert 0, '{} markers found!'.format(len(x))
marker_pos = np.asarray([x[0], y[0]])
reward = -1 * spatial.distance.cityblock(agent_pos[:2], marker_pos)
done = reward == 0
reward = float(done)
self.done = self.done or done
return reward, done
def _get_fourCorners_task_reward(self, agent_pos):
done = False
reward = 0
w = self.w
h = self.h
state = self.s_h[-1]
#give 0.25 reward for putting marker in each corner
if state[1, 1, 6]:
reward += 0.25
if state[h-2, 1, 6]:
reward += 0.25
if state[h-2, w-2, 6]:
reward += 0.25
if state[1, w-2, 6]:
reward += 0.25
#give zero reward if agent places marker anywhere else
correct_markers = int(reward*4)
incorrect_markers = self.total_markers - correct_markers
if incorrect_markers > 0:
reward = 0
done = reward == 1
reward = reward if self.env_task == 'fourCorners' else float(done)
if self.env_task == 'fourCorners_sparse':
reward = reward if done and not self.done else 0
self.done = self.done or done
return reward, done
def _get_stairClimber_task_reward(self, agent_pos):
done = False
reward = 0
state = self.s_h[-1]
# initial marker position
init_state = self.s_h[0]
x, y = np.where(init_state[:, :, 6] > 0)
if len(x) != 1: assert 0, '{} markers found!'.format(len(x))
marker_pos = np.asarray([x[0], y[0]])
reward = -1 * spatial.distance.cityblock(agent_pos[:2], marker_pos)
# NOTE: need to do this to avoid high negative reward for first action
if len(self.s_h) == 2:
x, y, z = np.where(self.s_h[0][:, :, :4] > 0)
prev_pos = np.asarray([x[0], y[0], z[0]])
self.prev_pos_reward = -1 * spatial.distance.cityblock(prev_pos[:2], marker_pos)
if not self.reward_diff:
# since reward is based on manhattan distance, rescale it to range between 0 to 1
if self.rescale_reward:
from_min, from_max, to_min, to_max = -(sum(self.s.shape[:2])), 0, -1, 0
reward = ((reward - from_min) * (to_max - to_min) / (from_max - from_min)) + to_min
if tuple(agent_pos[:2]) not in self.metadata['agent_valid_positions']:
reward = -1.0
done = reward == 0
else:
abs_reward = reward
reward = self.prev_pos_reward-1.0 if tuple(agent_pos[:2]) not in self.metadata['agent_valid_positions'] else reward
reward = reward - self.prev_pos_reward
self.prev_pos_reward = abs_reward
done = abs_reward == 0
reward = reward if self.env_task == 'stairClimber' else float(done)
if self.env_task == 'stairClimber_sparse':
reward = reward if done and not self.done else 0
self.done = self.done or done
return reward, done
def _get_topOff_task_reward(self, agent_pos):
assert self.reward_diff
done = False
reward = 0
w = self.w
h = self.h
state = self.s_h[-1]
for c in range(1, agent_pos[1]+1):
if (h-2, c) in self.metadata['not_expected_marker_positions']:
if state[h-2, c, 7]:
reward += 1
else:
break
else:
assert (h-2, c) in self.metadata['expected_marker_positions']
if state[h-2, c, 5]:
reward += 1
else:
break
if (self.w - 2 == agent_pos[1] and self.h - 2 == agent_pos[0]) and reward == w-2:
reward += 1
reward = reward / (w-1)
done = sum([state[pos[0], pos[1], 7] for pos in self.metadata['not_expected_marker_positions']]) == len(
self.metadata['not_expected_marker_positions']) and (self.w - 2 == agent_pos[1] and self.h - 2 == agent_pos[0]) and reward==1.0
reward = reward if self.env_task == 'topOff' else float(done)
if self.env_task == 'topOff_sparse':
reward = reward if done and not self.done else 0
self.done = self.done or done
return reward, done
def _get_state_reward(self, agent_pos, made_error=False):
if self.env_task == 'cleanHouse' or self.env_task == 'cleanHouse_sparse':
reward, done = self._get_cleanHouse_task_reward(agent_pos)
elif self.env_task == 'harvester' or self.env_task == 'harvester_sparse':
reward, done = self._get_harvester_task_reward(agent_pos)
elif self.env_task == 'fourCorners' or self.env_task == 'fourCorners_sparse':
reward, done = self._get_fourCorners_task_reward(agent_pos)
elif self.env_task == 'randomMaze' or self.env_task == 'randomMaze_sparse':
reward, done = self._get_randomMaze_task_reward(agent_pos)
elif self.env_task == 'stairClimber' or self.env_task == 'stairClimber_sparse':
reward, done = self._get_stairClimber_task_reward(agent_pos)
elif self.env_task == 'topOff' or self.env_task == 'topOff_sparse':
reward, done = self._get_topOff_task_reward(agent_pos)
else:
raise NotImplementedError('{} task not yet supported'.format(self.env_task))
return reward, done
def print_state(self, state=None):
agent_direction = {0: 'N', 1: 'E', 2: 'S', 3: 'W'}
state = self.s_h[-1] if state is None else state
state_2d = np.chararray(state.shape[:2])
state_2d[:] = '.'
state_2d[state[:,:,4]] = 'x'
state_2d[state[:,:,6]] = 'M'
x, y, z = np.where(state[:, :, :4] > 0)
state_2d[x[0], y[0]] = agent_direction[z[0]]
state_2d = state_2d.decode()
for i in range(state_2d.shape[0]):
print("".join(state_2d[i]))
def render(self, mode='rgb_array'):
return self.s_h[-1]
# get location (x, y) and facing {north, east, south, west}
def get_location(self):
x, y, z = np.where(self.s[:, :, :4] > 0)
return np.asarray([x[0], y[0], z[0]])
# get the neighbor {front, left, right} location
def get_neighbor(self, face):
loc = self.get_location()
if face == 'front':
neighbor_loc = loc[:2] + {
0: [-1, 0],
1: [0, 1],
2: [1, 0],
3: [0, -1]
}[loc[2]]
elif face == 'left':
neighbor_loc = loc[:2] + {
0: [0, -1],
1: [-1, 0],
2: [0, 1],
3: [1, 0]
}[loc[2]]
elif face == 'right':
neighbor_loc = loc[:2] + {
0: [0, 1],
1: [1, 0],
2: [0, -1],
3: [-1, 0]
}[loc[2]]
return neighbor_loc
###################################
### Perception Primitives ###
###################################
# return if the neighbor {front, left, right} of Karel is clear
def neighbor_is_clear(self, face):
self.perception_count += 1
neighbor_loc = self.get_neighbor(face)
if neighbor_loc[0] >= self.h or neighbor_loc[0] < 0 \
or neighbor_loc[1] >= self.w or neighbor_loc[1] < 0:
return False
return not self.s[neighbor_loc[0], neighbor_loc[1], 4]
def front_is_clear(self):
return self.neighbor_is_clear('front')
def left_is_clear(self):
return self.neighbor_is_clear('left')
def right_is_clear(self):
return self.neighbor_is_clear('right')
# return if there is a marker presented
def marker_present(self):
self.perception_count += 1
loc = self.get_location()
return np.sum(self.s[loc[0], loc[1], 6:]) > 0
def no_marker_present(self):
self.perception_count += 1
loc = self.get_location()
return np.sum(self.s[loc[0], loc[1], 6:]) == 0
def get_perception_list(self):
vec = ['frontIsClear', 'leftIsClear',
'rightIsClear', 'markersPresent',
'noMarkersPresent']
return vec
def get_perception_vector(self):
vec = [self.front_is_clear(), self.left_is_clear(),
self.right_is_clear(), self.marker_present(),
self.no_marker_present()]
return np.array(vec)
###################################
### State Transition ###
###################################
# given a state and a action, return the next state
def state_transition(self, a):
made_error = False
a_idx = np.argmax(a)
loc = self.get_location()
if a_idx == 0:
# move
if self.front_is_clear():
front_loc = self.get_neighbor('front')
loc_vec = self.s[loc[0], loc[1], :4]
self.s[front_loc[0], front_loc[1], :4] = loc_vec
self.s[loc[0], loc[1], :4] = np.zeros(4) > 0
next_loc = front_loc
else:
if self.make_error:
raise RuntimeError("Failed to move.")
loc_vec = np.zeros(4) > 0
loc_vec[(loc[2] + 2) % 4] = True # Turn 180
self.s[loc[0], loc[1], :4] = loc_vec
next_loc = loc
self.add_to_history(a_idx, next_loc)
elif a_idx == 1 or a_idx == 2:
# turn left or right
loc_vec = np.zeros(4) > 0
loc_vec[(a_idx * 2 - 3 + loc[2]) % 4] = True
self.s[loc[0], loc[1], :4] = loc_vec
self.add_to_history(a_idx, loc)
elif a_idx == 3 or a_idx == 4:
# pick up or put a marker
num_marker = np.argmax(self.s[loc[0], loc[1], 5:])
# just clip the num of markers for now
# new_num_marker = np.clip(a_idx*2-7 + num_marker, 0, MAX_NUM_MARKER-1)
new_num_marker = a_idx*2-7 + num_marker
if new_num_marker < 0:
if self.make_error:
raise RuntimeError("No marker to pick up.")
else:
new_num_marker = num_marker
made_error = True
elif new_num_marker > MAX_NUM_MARKER-1:
if self.make_error:
raise RuntimeError("Cannot put more marker.")
else:
new_num_marker = num_marker
made_error = True
marker_vec = np.zeros(MAX_NUM_MARKER+1) > 0
marker_vec[new_num_marker] = True
self.s[loc[0], loc[1], 5:] = marker_vec
self.add_to_history(a_idx, loc, made_error)
else:
raise RuntimeError("Invalid action")
return
# given a karel env state, return a visulized image
def state2image(self, s=None, grid_size=100, root_dir='./'):
h = s.shape[0]
w = s.shape[1]
img = np.ones((h*grid_size, w*grid_size, 1))
import pickle
from PIL import Image
import os.path as osp
f = pickle.load(open(osp.join(root_dir, 'karel_env/asset/texture.pkl'), 'rb'))
wall_img = f['wall'].astype('uint8')
marker_img = f['marker'].astype('uint8')
agent_0_img = f['agent_0'].astype('uint8')
agent_1_img = f['agent_1'].astype('uint8')
agent_2_img = f['agent_2'].astype('uint8')
agent_3_img = f['agent_3'].astype('uint8')
blank_img = f['blank'].astype('uint8')
#blanks
for y in range(h):
for x in range(w):
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = blank_img
# wall
y, x = np.where(s[:, :, 4])
for i in range(len(x)):
img[y[i]*grid_size:(y[i]+1)*grid_size, x[i]*grid_size:(x[i]+1)*grid_size] = wall_img
# marker
y, x = np.where(np.sum(s[:, :, 6:], axis=-1))
for i in range(len(x)):
img[y[i]*grid_size:(y[i]+1)*grid_size, x[i]*grid_size:(x[i]+1)*grid_size] = marker_img
# karel
y, x = np.where(np.sum(s[:, :, :4], axis=-1))
if len(y) == 1:
y = y[0]
x = x[0]
idx = np.argmax(s[y, x])
marker_present = np.sum(s[y, x, 6:]) > 0
if marker_present:
extra_marker_img = Image.fromarray(f['marker'].squeeze()).copy()
if idx == 0:
extra_marker_img.paste(Image.fromarray(f['agent_0'].squeeze()))
extra_marker_img = f['marker'].squeeze() + f['agent_0'].squeeze()
extra_marker_img = np.minimum(f['marker'].squeeze() , f['agent_0'].squeeze())
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = np.expand_dims(np.array(extra_marker_img), axis=-1)
elif idx == 1:
extra_marker_img.paste(Image.fromarray(f['agent_1'].squeeze()))
extra_marker_img = f['marker'].squeeze() + f['agent_1'].squeeze()
extra_marker_img = np.minimum(f['marker'].squeeze() , f['agent_1'].squeeze())
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = np.expand_dims(np.array(extra_marker_img), axis=-1)
elif idx == 2:
extra_marker_img.paste(Image.fromarray(f['agent_2'].squeeze()))
extra_marker_img = f['marker'].squeeze() + f['agent_2'].squeeze()
extra_marker_img = np.minimum(f['marker'].squeeze() , f['agent_2'].squeeze())
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = np.expand_dims(np.array(extra_marker_img), axis=-1)
elif idx == 3:
extra_marker_img.paste(Image.fromarray(f['agent_3'].squeeze()))
extra_marker_img = f['marker'].squeeze() + f['agent_3'].squeeze()
extra_marker_img = np.minimum(f['marker'].squeeze() , f['agent_3'].squeeze())
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = np.expand_dims(np.array(extra_marker_img), axis=-1)
else:
if idx == 0:
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = f['agent_0']
elif idx == 1:
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = f['agent_1']
elif idx == 2:
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = f['agent_2']
elif idx == 3:
img[y*grid_size:(y+1)*grid_size, x*grid_size:(x+1)*grid_size] = f['agent_3']
elif len(y) > 1:
raise ValueError
return img