-
Notifications
You must be signed in to change notification settings - Fork 2
/
demo.py
167 lines (155 loc) · 6.35 KB
/
demo.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
# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
#
# Author: Wentao Yuan
'''
Demo script that shows data loading and model inference.
'''
import hydra
import numpy as np
import torch
from m2t2.dataset import load_rgb_xyz, collate
from m2t2.dataset_utils import denormalize_rgb, sample_points
from m2t2.meshcat_utils import (
create_visualizer, make_frame, visualize_grasp, visualize_pointcloud
)
from m2t2.m2t2 import M2T2
from m2t2.plot_utils import get_set_colors
from m2t2.train_utils import to_cpu, to_gpu
def load_and_predict(data_dir, cfg):
data, meta_data = load_rgb_xyz(
data_dir, cfg.data.robot_prob,
cfg.data.world_coord, cfg.data.jitter_scale,
cfg.data.grid_resolution, cfg.eval.surface_range
)
if 'object_label' in meta_data:
data['task'] = 'place'
else:
data['task'] = 'pick'
model = M2T2.from_config(cfg.m2t2)
ckpt = torch.load(cfg.eval.checkpoint)
model.load_state_dict(ckpt['model'])
model = model.cuda().eval()
inputs, xyz, seg = data['inputs'], data['points'], data['seg']
obj_inputs = data['object_inputs']
outputs = {
'grasps': [],
'grasp_confidence': [],
'grasp_contacts': [],
'placements': [],
'placement_confidence': [],
'placement_contacts': []
}
for _ in range(cfg.eval.num_runs):
pt_idx = sample_points(xyz, cfg.data.num_points)
data['inputs'] = inputs[pt_idx]
data['points'] = xyz[pt_idx]
data['seg'] = seg[pt_idx]
pt_idx = sample_points(obj_inputs, cfg.data.num_object_points)
data['object_inputs'] = obj_inputs[pt_idx]
data_batch = collate([data])
to_gpu(data_batch)
with torch.no_grad():
model_ouputs = model.infer(data_batch, cfg.eval)
to_cpu(model_ouputs)
for key in outputs:
if 'place' in key and len(outputs[key]) > 0:
outputs[key] = [
torch.cat([prev, cur])
for prev, cur in zip(outputs[key], model_ouputs[key][0])
]
else:
outputs[key].extend(model_ouputs[key][0])
data['inputs'], data['points'], data['seg'] = inputs, xyz, seg
data['object_inputs'] = obj_inputs
return data, outputs
@hydra.main(config_path='.', config_name='config', version_base='1.3')
def main(cfg):
data, outputs = load_and_predict(cfg.eval.data_dir, cfg)
vis = create_visualizer()
rgb = denormalize_rgb(
data['inputs'][:, 3:].T.unsqueeze(2)
).squeeze(2).T
rgb = (rgb.numpy() * 255).astype('uint8')
xyz = data['points'].numpy()
cam_pose = data['cam_pose'].double().numpy()
make_frame(vis, 'camera', T=cam_pose)
if not cfg.eval.world_coord:
xyz = xyz @ cam_pose[:3, :3].T + cam_pose[:3, 3]
visualize_pointcloud(vis, 'scene', xyz, rgb, size=0.005)
if data['task'] == 'pick':
for i, (grasps, conf, contacts, color) in enumerate(zip(
outputs['grasps'],
outputs['grasp_confidence'],
outputs['grasp_contacts'],
get_set_colors()
)):
print(f"object_{i:02d} has {grasps.shape[0]} grasps")
conf = conf.numpy()
conf_colors = (np.stack([
1 - conf, conf, np.zeros_like(conf)
], axis=1) * 255).astype('uint8')
visualize_pointcloud(
vis, f"object_{i:02d}/contacts",
contacts.numpy(), conf_colors, size=0.01
)
grasps = grasps.numpy()
if not cfg.eval.world_coord:
grasps = cam_pose @ grasps
for j, grasp in enumerate(grasps):
visualize_grasp(
vis, f"object_{i:02d}/grasps/{j:03d}",
grasp, color, linewidth=0.2
)
elif data['task'] == 'place':
ee_pose = data['ee_pose'].double().numpy()
make_frame(vis, 'ee', T=ee_pose)
obj_xyz_ee, obj_rgb = data['object_inputs'].split([3, 3], dim=1)
obj_xyz_ee = (obj_xyz_ee + data['object_center']).numpy()
obj_xyz = obj_xyz_ee @ ee_pose[:3, :3].T + ee_pose[:3, 3]
obj_rgb = denormalize_rgb(obj_rgb.T.unsqueeze(2)).squeeze(2).T
obj_rgb = (obj_rgb.numpy() * 255).astype('uint8')
visualize_pointcloud(vis, 'object', obj_xyz, obj_rgb, size=0.005)
for i, (placements, conf, contacts) in enumerate(zip(
outputs['placements'],
outputs['placement_confidence'],
outputs['placement_contacts'],
)):
print(f"orientation_{i:02d} has {placements.shape[0]} placements")
conf = conf.numpy()
conf_colors = (np.stack([
1 - conf, conf, np.zeros_like(conf)
], axis=1) * 255).astype('uint8')
visualize_pointcloud(
vis, f"orientation_{i:02d}/contacts",
contacts.numpy(), conf_colors, size=0.01
)
placements = placements.numpy()
if not cfg.eval.world_coord:
placements = cam_pose @ placements
visited = np.zeros((0, 3))
for j, k in enumerate(np.random.permutation(placements.shape[0])):
if visited.shape[0] > 0:
dist = np.sqrt((
(placements[k, :3, 3] - visited) ** 2
).sum(axis=1))
if dist.min() < cfg.eval.placement_vis_radius:
continue
visited = np.concatenate([visited, placements[k:k+1, :3, 3]])
visualize_grasp(
vis, f"orientation_{i:02d}/placements/{j:02d}/gripper",
placements[k], [0, 255, 0], linewidth=0.2
)
obj_xyz_placed = obj_xyz_ee @ placements[k, :3, :3].T \
+ placements[k, :3, 3]
visualize_pointcloud(
vis, f"orientation_{i:02d}/placements/{j:02d}/object",
obj_xyz_placed, obj_rgb, size=0.01
)
if __name__ == '__main__':
main()