/
collect_demos.py
156 lines (147 loc) · 6.24 KB
/
collect_demos.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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import torch
import numpy as np
import gym
import envs
from evaluate import preproc_inputs
ENV_2_CKPT_RL = {
'Fetch3Push-v1': './weights/3p-3uhwocsy.pt',
'Fetch3Switch-v1': './weights/3s-2mb98s47.pt',
'Fetch2Switch2Push-v1': './weights/2s2p-338rpvyu.pt',
}
@torch.no_grad()
def main(args):
if args.chain:
if args.env in ['FetchStack2Stage3-v1', 'FetchStack2StitchOnlyStack-v1', 'FetchStack2Stage1-v1']:
model_paths = {
'push': './weights/stack-1p-1i0ac7fq.pt',
'stack': './weights/stack-1s-2q3z23zd.pt',
}
networks = {}
for name, model_path in model_paths.items():
x_norm, actor_network = torch.load(model_path, map_location=lambda storage, loc: storage)
actor_network.eval()
networks[name] = (x_norm, actor_network)
def policy(observation):
next_obj_idx = observation['next_object_idx']
if args.env == 'FetchStack2Stage1-v1':
name = 'push'
else:
name = 'push' if next_obj_idx == 0 else 'stack'
grip = observation['gripper_arr']
obj = observation['object_arr']
g = observation['desired_goal_arr']
x_norm, actor_network = networks[name]
inputs = preproc_inputs(grip, obj, g, x_norm)
pi = actor_network(*inputs)
return pi.numpy().squeeze()
else:
model_paths = {}
if 'Push' in args.env:
model_paths['push'] = './weights/1p-3n2mu1hy.pt'
if 'Switch' in args.env:
model_paths['switch'] = './weights/1s-1gde5bpj.pt'
networks = {}
for name, model_path in model_paths.items():
x_norm, actor_network = torch.load(model_path, map_location=lambda storage, loc: storage)
actor_network.eval()
networks[name] = (x_norm, actor_network)
def policy(observation):
next_obj_idx = observation['next_object_idx']
grip = observation['gripper_arr'][None]
obj = observation['object_arr'][next_obj_idx][None]
obj_type = np.squeeze(obj[..., -1])
name = 'switch' if obj_type > 0 else 'push'
g = observation['desired_goal_arr'][next_obj_idx][None]
obj, g = obj[None], g[None]
x_norm, actor_network = networks[name]
inputs = preproc_inputs(grip, obj, g, x_norm)
pi = actor_network(*inputs)
return pi.numpy().squeeze()
else:
model_path = ENV_2_CKPT_RL[args.env]
x_norm, actor_network = torch.load(model_path, map_location=lambda storage, loc: storage)
actor_network.eval()
def policy(observation):
grip = observation['gripper_arr'][None]
obj, g = observation['object_arr'], observation['desired_goal_arr']
obj, g = obj[None], g[None]
inputs = preproc_inputs(grip, obj, g, x_norm)
pi = actor_network(*inputs)
return pi.numpy().squeeze()
print(f"Collecting demos in {args.env}.")
env = gym.make(args.env)
observation = env.reset()
if args.render:
env = gym.wrappers.Monitor(env, "./demo_videos", force=True)
observation = env.reset()
successes = []
ret_arr, solved_t_arr = [], []
grip_arr, obj_arr, act_arr, ag_arr, goal_arr, success_arr = [], [], [], [], [], []
for i in range(args.num_eps):
# start to do the demo
ep_grip, ep_obj, ep_act, ep_ag, ep_g, ep_success = [], [], [], [], [], []
observation = env.reset()
ret, solved_t = 0, -1
for t in range(env._max_episode_steps):
action = policy(observation)
# put actions into the environment
observation_new, rew, done, info = env.step(action)
ep_grip.append(observation["gripper_arr"].copy())
ep_obj.append(observation["object_arr"].copy())
ep_ag.append(observation["achieved_goal_arr"].copy())
ep_g.append(observation["desired_goal_arr"].copy())
ep_act.append(action.copy())
ep_success.append(info["is_success"])
observation = observation_new
if solved_t < 0:
ret += rew
if info['is_success'] and solved_t < 0:
solved_t = t
ep_grip.append(observation["gripper_arr"].copy())
ep_obj.append(observation["object_arr"].copy())
ep_ag.append(observation["achieved_goal_arr"].copy())
print(f'episode {i}, is success: {info["is_success"]}, finished t: {solved_t}')
successes.append(info['is_success'])
grip_arr.append(np.stack(ep_grip))
obj_arr.append(np.stack(ep_obj))
act_arr.append(np.stack(ep_act))
ag_arr.append(np.stack(ep_ag))
goal_arr.append(np.stack(ep_g))
success_arr.append(np.stack(ep_success))
ret_arr.append(ret)
solved_t_arr.append(env._max_episode_steps if solved_t < 0 else solved_t)
grip_arr = np.stack(grip_arr)
obj_arr = np.stack(obj_arr)
act_arr = np.stack(act_arr)
ag_arr = np.stack(ag_arr)
goal_arr = np.stack(goal_arr)
success_arr = np.stack(success_arr)
ret_arr = np.array(ret_arr)
solved_t_arr = np.array(solved_t_arr)
name = "init_trajs" if args.chain else "rl_expert"
np.savez(
f"./data/{name}_{args.env}_{args.num_eps}.npz",
grip=grip_arr,
obj=obj_arr,
action=act_arr,
ag=ag_arr,
g=goal_arr,
success=success_arr,
ret_arr=ret_arr,
solved_t_arr=solved_t_arr,
)
print(np.mean(successes))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("env", type=str, default="Fetch3Push-v1")
parser.add_argument("--chain", action="store_true")
parser.add_argument("--num_eps", type=int, default=1500)
parser.add_argument("--render", action="store_true")
args = parser.parse_args()
main(args)