コンテンツにスキップ

HilSerl 環境構築

HilSerl公式

https://hil-serl.github.io/

検証環境

  • RTX5090
  • Ubuntu24.04
  • Anaconda

Anaconda環境の作成

1
2
conda create -y -n lerobot_rl python=3.10
conda activate lerobot_rl

1
conda install ffmpeg -c conda-forge

RTX5090は、CUDA12.8対応なので、nightly buildから展開

1
2
pip install --extra-index-url https://download.pytorch.org/whl/nightly/cu128 \
torch==2.8.0.dev20250603+cu128 torchvision==0.23.0.dev20250603+cu128 torchaudio==2.8.0.dev20250603+cu128

1
pip install triton==2.1.0 --index-url https://download.pytorch.org/whl/nightly/cu128

チェック

1
2
python -c "import torch; print(torch.__version__)"
2.8.0.dev20250603+cu128

1
2
python -c "import torchvision; print(torchvision.__version__)"
0.23.0.dev20250603+cu128

1
2
python -c "import torchaudio; print(torchaudio.__version__)"
2.8.0.dev20250603+cu128

Transformersもインストール

1
conda install -c conda-forge transformers accelerate safetensors

LeRobotをGit clone

1
git clone https://github.com/huggingface/lerobot/ lerobot_rl

1
cd lerobot_rl

pyproject.toml

 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
dependencies = [
    "cmake>=3.29.0.1",
    "datasets>=2.19.0",
    "deepdiff>=7.0.1",
    "diffusers>=0.27.2",
    "draccus==0.10.0",
    "einops>=0.8.0",
    "flask>=3.0.3",
    "gdown>=5.1.0",
    "gymnasium==0.29.1",                                                 # TODO(rcadene, aliberts): Make gym 1.0.0 work
    "h5py>=3.10.0",
    "huggingface-hub[hf-transfer,cli]>=0.27.1 ; python_version < '4.0'",
    "imageio[ffmpeg]>=2.34.0",
    "jsonlines>=4.0.0",
    "numba>=0.59.0",
    "omegaconf>=2.3.0",
    "opencv-python-headless>=4.9.0",
    "packaging>=24.2",
    "av>=14.2.0",
    "pymunk>=6.6.0",
    "pynput>=1.7.7",
    "pyzmq>=26.2.1",
    "rerun-sdk>=0.21.0",
    "termcolor>=2.4.0",
    #"torch>=2.2.1,<2.7",
    #"torchcodec==0.2.1; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
    #"torchvision>=0.21.0",
    "wandb>=0.16.3",
    "zarr>=2.17.0",
]

1
2
pip install -e ".[feetech]"
pip install -e ".[hilserl]"

1
pip install grpcio grpcio-tools wandb gymnasium mujoco mujoco-py

1
pip install scipy pygame

各種パラメーター解説

パラメーター 説明
end_effector_step_sizes 1ステップあたり最大移動量[cm]を設定
end_effector_bounds min：エンドエフェクタ位置の XYZ 各軸の最小値,max：エンドエフェクタ位置の XYZ 各軸の最大値
root/type hilとgym_manipulatorが設定可能
robot/type so100_follower_end_effector, so101_follower_end_effector, koch_follower_end_effector, moss_follower_end_effector
teleop/type so101_leader, so100_leader, koch_leader, gamepad, widowx, stretch3_gamepad, keyboard

カメラ設定例

Realsense

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
    "type": "gym_manipulator",
    "robot": {
        "cameras": {
          "front": {
            "type": "intelrealsense",
            "serial_number_or_name": "008222070291",
            "fps": 30,
            "width": 640,
            "height": 480
            },
          "wrist": {
            "type": "intelrealsense",
            "serial_number_or_name": "138422077679",
            "fps": 30,
            "width": 640,
            "height": 480
            }
        }
    },

USB Camera

1
v4l2-ctl -d /dev/video0 --list-formats-ext

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
    "type": "gym_manipulator",
    "robot": {
        "cameras": {
            "front": {
                "type": "opencv",
                "index_or_path": 0,
                "height": 480,
                "width": 640,
                "fps": 30
            },
            "wrist": {
                "type": "opencv",
                "index_or_path": 2,
                "height": 480,
                "width": 640,
                "fps": 30
            }
        }
    },

URDF

1
2
cd ~
git clone https://github.com/TheRobotStudio/SO-ARM100

1
2
3
4
cd SO-ARM100/Simulation/SO101/
(base) ubuntu@rtx5090:~/SO-ARM100/Simulation/SO101$ ls
assets                 README.md  so101_new_calib.urdf  so101_old_calib.urdf
joints_properties.xml  scene.xml  so101_new_calib.xml   so101_old_calib.xml

so101_new_calib.urdfのフォルダをメモし、configに反映

end_effector_boundsの値の取得

1
2
3
4
5
6
7
8
python lerobot/scripts/find_joint_limits.py  \
    --robot.type=so100_follower_end_effector  \
    --robot.port=/dev/ttyACM1  \
    --robot.id=black  \
    --teleop.type=so100_leader  \
    --teleop.port=/dev/ttyACM0  \
    --teleop.id=blue \
    --robot.urdf_path=/home/ubuntu/SO-ARM100/Simulation/SO101/so101_new_calib.urdf

うまく動かない。原因調査中。

SO101アーム使用する場合のConfig

Spaceを押すと手動モード

Info

JSONファイルの設定は、aractingi/lerobot-example-config-files　を参考に

  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
{
    "type": "gym_manipulator",
    "robot": {
        "type": "so100_follower_end_effector",
        "port": "/dev/ttyACM1",
        "urdf_path": "/home/ubuntu/SO-ARM100/Simulation/SO101/so101_new_calib.urdf",
        "target_frame_name": "gripper_frame_link",
        "cameras": {
            "front": {
                "type": "opencv",
                "index_or_path": 0,
                "height": 480,
                "width": 640,
                "fps": 30
            },
            "wrist": {
                "type": "opencv",
                "index_or_path": 2,
                "height": 480,
                "width": 640,
                "fps": 30
            }
        },
        "end_effector_bounds": {
            "min": [
                -1.0,
                -1.0,
                -1.0
            ],
            "max": [
                1.0,
                1.0,
                1.0
            ]
        },
        "end_effector_step_sizes": {
            "x": 0.01,
            "y": 0.01,
            "z": 0.01
        }
    },
    "teleop": {
      "type": "so101_leader",
      "port": "/dev/ttyACM0",
      "use_degrees": true
    },
    "wrapper": {
        "display_cameras": false,
        "add_joint_velocity_to_observation": true,
        "add_current_to_observation": true,
        "add_ee_pose_to_observation": true,
        "crop_params_dict": {
            "observation.images.front": [
                270,
                170,
                90,
                190
            ],
            "observation.images.wrist": [
                0,
                0,
                480,
                640
            ]
        },
        "resize_size": [
            128,
            128
        ],
        "control_time_s": 20.0,
        "use_gripper": true,
        "gripper_quantization_threshold": null,
        "gripper_penalty": -0.02,
        "gripper_penalty_in_reward": false,
        "fixed_reset_joint_positions": [
            0.0,
            0.0,
            0.0,
            90.0,
            0.0,
            5.0
        ],
        "reset_time_s": 2.5,
        "control_mode": "leader"
    },
    "name": "real_robot",
    "mode": "record",
    "repo_id": "akira-sasaki/hilserl_test",
    "dataset_root": null,
    "task": "",
    "num_episodes": 2,
    "episode": 0,
    "pretrained_policy_name_or_path": null,
    "device": "cpu",
    "push_to_hub": true,
    "fps": 10,
    "features": {
        "observation.images.front": {
            "type": "VISUAL",
            "shape": [
                3,
                128,
                128
            ]
        },
        "observation.images.wrist": {
            "type": "VISUAL",
            "shape": [
                3,
                128,
                128
            ]
        },
        "observation.state": {
            "type": "STATE",
            "shape": [
                15
            ]
        },
        "action": {
            "type": "ACTION",
            "shape": [
                3
            ]
        }
    },
    "features_map": {
        "observation.images.front": "observation.images.side",
        "observation.images.wrist": "observation.images.wrist",
        "observation.state": "observation.state",
        "action": "action"
    },
    "reward_classifier_pretrained_path": null
}

初期学習

1
2
python lerobot/scripts/rl/gym_manipulator.py \
    --config_path lerobot/configs/env_config_so100.json

calibrationファイルは、

デバイス名 ファイルのDIR
so100_follower_end_effector ~/.cache/huggingface/lerobot/calibration/robots/so100_follower_end_effector
so100_leader ~/.cache/huggingface/lerobot/calibration/teleoperators/so100_leader

Train_config

  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
{
    "output_dir": null,
    "job_name": "default",
    "resume": false,
    "seed": 1000,
    "num_workers": 4,
    "batch_size": 256,
    "steps": 100000,
    "log_freq": 500,
    "save_checkpoint": true,
    "save_freq": 2000000,
    "wandb": {
        "enable": true,
        "project": "so100_real",
        "disable_artifact": true
    },
    "dataset": {
        "repo_id": "akira-sasaki/hilserl_test",
        "use_imagenet_stats": false
    },
    "policy": {
        "type": "sac",
        "n_obs_steps": 1,
        "repo_id": "akira-sasaki/policy_hilserl_test",
        "normalization_mapping": {
            "VISUAL": "MEAN_STD",
            "STATE": "MIN_MAX",
            "ENV": "MIN_MAX",
            "ACTION": "MIN_MAX"
        },
        "input_features": {
            "observation.images.side": {
                "type": "VISUAL",
                "shape": [
                    3,
                    128,
                    128
                ]
            },
            "observation.images.front": {
                "type": "VISUAL",
                "shape": [
                    3,
                    128,
                    128
                ]
            },
            "observation.state": {
                "type": "STATE",
                "shape": [
                    15
                ]
            }
        },
        "device": "cuda",
        "use_amp": false,
        "dataset_stats": {
            "observation.images.side": {
                "mean": [
                    0.485,
                    0.456,
                    0.406
                ],
                "std": [
                    0.229,
                    0.224,
                    0.225
                ]
            },
            "observation.images.front": {
                "mean": [
                    0.485,
                    0.456,
                    0.406
                ],
                "std": [
                    0.229,
                    0.224,
                    0.225
                ]
            },
            "observation.state": {
                "min": [
                    -15.0,
                    -20.0,
                    10.0,
                    65.0,
                    0.0,
                    0.0,
                    -40.0,
                    -60.0,
                    -50.0,
                    -80.0,
                    -10.0,
                    -110.0,
                    -15.0,
                    -20.0,
                    15.0,
                    70.0,
                    -5.0,
                    0.0,
                    0.26,
                    -0.06,
                    0.25
                ],
                "max": [
                    15.0,
                    15.0,
                    40.0,
                    110.0,
                    0.5,
                    31.0,
                    53.0,
                    61.0,
                    67.0,
                    90.0,
                    6.0,
                    100.0,
                    15.0,
                    12.0,
                    40.0,
                    110.0,
                    0.5,
                    31.0,
                    0.32,
                    0.06,
                    0.35
                ]
            },
            "action": {
                "min": [
                    -0.03,
                    -0.03,
                    -0.03
                ],
                "max": [
                    0.03,
                    0.03,
                    0.03
                ]
            }
        },
        "num_discrete_actions": 3,
        "storage_device": "cuda",
        "vision_encoder_name": "helper2424/resnet10",
        "freeze_vision_encoder": true,
        "image_encoder_hidden_dim": 32,
        "shared_encoder": true,
        "online_steps": 1000000,
        "online_env_seed": 10000,
        "online_buffer_capacity": 30000,
        "offline_buffer_capacity": 10000,
        "online_step_before_learning": 100,
        "policy_update_freq": 1,
        "discount": 0.97,
        "async_prefetch": false,
        "temperature_init": 0.01,
        "num_critics": 2,
        "num_subsample_critics": null,
        "critic_lr": 0.0003,
        "actor_lr": 0.0003,
        "temperature_lr": 0.0003,
        "critic_target_update_weight": 0.005,
        "utd_ratio": 2,
        "state_encoder_hidden_dim": 256,
        "latent_dim": 256,
        "target_entropy": null,
        "use_backup_entropy": true,
        "grad_clip_norm": 40.0,
        "critic_network_kwargs": {
            "hidden_dims": [
                256,
                256
            ],
            "activate_final": true,
            "final_activation": null
        },
        "actor_network_kwargs": {
            "hidden_dims": [
                256,
                256
            ],
            "activate_final": true
        },
        "policy_kwargs": {
            "use_tanh_squash": true,
            "std_min": -5,
            "std_max": 2,
            "init_final": 0.05
        },
        "actor_learner_config": {
            "learner_host": "127.0.0.1",
            "learner_port": 50051,
            "policy_parameters_push_frequency": 4
        },
        "concurrency": {
            "actor": "threads",
            "learner": "threads"
        }
    },
    "env": {
        "type": "gym_manipulator",
        "robot": {
            "type": "so100_follower_end_effector",
            "port": "/dev/ttyACM1",
            "urdf_path": "/home/ubuntu/SO-ARM100/Simulation/SO101/so101_new_calib.urdf",
            "target_frame_name": "gripper_frame_link",
            "cameras": {
                "front": {
                    "type": "opencv",
                    "index_or_path": 2,
                    "height": 480,
                    "width": 640,
                    "fps": 30
                },
                "wrist": {
                    "type": "opencv",
                    "index_or_path": 8,
                    "height": 480,
                    "width": 640,
                    "fps": 30
                }
            },
            "end_effector_bounds": {
                "min": [
                    -1.0,
                    -1.0,
                    -1.0
                ],
                "max": [
                    1.0,
                    1.0,
                    1.0
                ]
            },
            "max_gripper_pos": 30
        },
        "teleop": {
          "type": "so101_leader",
          "port": "/dev/ttyACM0",
          "use_degrees": true
        },
        "wrapper": {
            "display_cameras": false,
            "add_joint_velocity_to_observation": true,
            "add_current_to_observation": true,
            "add_ee_pose_to_observation": true,
            "crop_params_dict": {
                "observation.images.front": [
                    270,
                    170,
                    90,
                    190
                ],
                "observation.images.wrist": [
                    0,
                    0,
                    480,
                    640
                ]
            },
            "resize_size": [
                128,
                128
            ],
            "control_time_s": 20.0,
            "use_gripper": true,
            "gripper_quantization_threshold": null,
            "gripper_penalty": -0.02,
            "gripper_penalty_in_reward": false,
            "fixed_reset_joint_positions": [
                0.0,
                -20.0,
                20.0,
                90.0,
                0.0,
                30.0
            ],
            "reset_time_s": 2.5,
            "control_mode": "leader"
        },
        "name": "real_robot",
        "mode": null,
        "repo_id": null,
        "dataset_root": null,
        "task": "",
        "num_episodes": 0,
        "episode": 0,
        "pretrained_policy_name_or_path": null,
        "device": "cuda",
        "push_to_hub": true,
        "fps": 10,
        "features": {
            "observation.images.front": {
                "type": "VISUAL",
                "shape": [
                    3,
                    128,
                    128
                ]
            },
            "observation.images.wrist": {
                "type": "VISUAL",
                "shape": [
                    3,
                    128,
                    128
                ]
            },
            "observation.state": {
                "type": "STATE",
                "shape": [
                    15
                ]
            },
            "action": {
                "type": "ACTION",
                "shape": [
                    3
                ]
            }
        },
        "features_map": {
            "observation.images.front": "observation.images.side",
            "observation.images.wrist": "observation.images.wrist",
            "observation.state": "observation.state",
            "action": "action"
        }
    }
}

Learner

1
2
pip install --upgrade --extra-index-url https://download.pytorch.org/whl/nightly \
    "triton>=3.3.0"

1
2
python lerobot/scripts/rl/learner_service.py \
    --config_path lerobot/configs/train_config_hilserl_so100.json

Actor

1
2
python lerobot/scripts/rl/actor.py \
    --config_path src/lerobot/configs/train_config_hilserl_so100.json

Reference