ロボットアーム(SO-ARM101)で、アヒルを掴んでみました 〜模倣学習〜

$ uv init --python 3.12 lerobot-workspace
$ cd lerobot-workspace
$ uv add "lerobot[feetech]>=0.5.1" matplotlib

// LeRobotの確認
$ uv run lerobot-info
- LeRobot version: 0.5.1
- Platform: macOS-15.7.3-arm64-arm-64bit
- Python version: 3.12.2
- Huggingface Hub version: 1.13.0
- Transformers version: N/A
- Datasets version: 4.8.5
- Numpy version: 2.2.6
- FFmpeg version: N/A
- PyTorch version: 2.10.0
- Is PyTorch built with CUDA support?: False
- Cuda version: N/A
- GPU model: N/A
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-can', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train', 'lerobot-train-tokenizer']

// Feetech SDK の確認
$ uv run python -c "import scservo_sdk; print('OK')"

// matplotlib の確認
$ uv run python -c "import matplotlib; print(matplotlib.__version__)"
3.10.9

// MPS の確認
$ uv run python -c "import torch; print(f'MPS: {torch.backends.mps.is_available()}')"
MPS: True

// PyTorch からの MPS利用 確認
uv run python -c "import torch; print(f'MPS available: {torch.backends.mps.is_available()}'); print(f'MPS built:{torch.backends.mps.is_built()}')"
MPS available: True
MPS built:True

$ uv run python preview_camera.py 1

$ bash record.sh

#!/bin/bash
# Main recording: 30 episodes of "Pick up the yellow duck" task

uv run lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem5B3D0430421 \
    --robot.id=my_follower \
    --robot.cameras="{ front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 15} }" \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem5B3D0413001 \
    --teleop.id=my_leader \
    --display_data=true \
    --dataset.repo_id=duck_pickup_v1 \
    --dataset.num_episodes=30 \
    --dataset.single_task="Pick up the yellow duck and put it in the basket" \
    --dataset.push_to_hub=false \
    --dataset.fps=15 \
    --dataset.episode_time_s=20 \
    --dataset.reset_time_s=10 \
    --dataset.video_encoding_batch_size=1 \
    --dataset.vcodec=h264_videotoolbox

$ tree .
.
├── data
│   └── chunk-000
│       └── file-000.parquet
├── meta
│   ├── episodes
│   │   └── chunk-000
│   │       └── file-000.parquet
│   ├── info.json
│   ├── stats.json
│   └── tasks.parquet
└── videos
    └── observation.images.front
        └── chunk-000
            ├── file-000.mp4
            └── file-001.mp4

$ caffeinate -d -i bash train_act.sh

uv run lerobot-train \
    --dataset.repo_id=duck_pickup_v1 \
    --dataset.video_backend=pyav \
    --policy.type=act \
    --policy.device=mps \
    --output_dir=outputs/train/act_duck_pickup_v1 \
    --steps=30000 \
    --save_freq=5000 \
    --wandb.enable=false \
    --policy.push_to_hub=false

sudo powermetrics --samplers gpu_power -i 1000

**** GPU usage ****

GPU HW active frequency: 1398 MHz
GPU HW active residency:  99.78% (444 MHz: .01% 612 MHz: 0% 808 MHz: 0%
                                  968 MHz:   0% 1110 MHz: 0% 1236 MHz: 0%
                                  1338 MHz:   0% 1398 MHz: 100%)
GPU SW requested state: (P1 : 0% P2 : 0% P3 : 0% P4 : 0% P5 : 0% P6 : 0% P7 : 0% P8 : 100%)
GPU idle residency:   0.22%
GPU Power: 9796 mW

**** GPU usage ****

GPU HW active frequency: 444 MHz
GPU HW active residency:  48.02% (444 MHz:  48% 612 MHz:   0% 808 MHz:   0% 968 MHz:   0% 1110 MHz:   0% 1236 MHz:   0% 1338 MHz:   0% 1398 MHz:   0%)
GPU SW requested state: (P1 : 100% P2 :   0% P3 :   0% P4 :   0% P5 :   0% P6 :   0% P7 :   0% P8 :   0%)
GPU idle residency:  51.98%
GPU Power: 225 mW

$ bash eval_act.sh

#!/bin/bash
# Evaluate trained ACT model - 1 episode only (records video)

#CHECKPOINT="last"
CHECKPOINT="025000"
EVAL_NAME="eval_${CHECKPOINT}_single_$(date +%Y%m%d_%H%M%S)"

uv run lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem5B3D0430421 \
    --robot.id=my_follower \
    --robot.cameras="{ front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 15} }" \
    --display_data=true \
    --dataset.repo_id=${EVAL_NAME} \
    --dataset.num_episodes=1 \
    --dataset.single_task="Pick up the yellow duck and put it in the basket" \
    --dataset.push_to_hub=false \
    --dataset.episode_time_s=20 \
    --dataset.reset_time_s=5 \
    --dataset.fps=15 \
    --dataset.video_encoding_batch_size=1 \
    --dataset.vcodec=h264_videotoolbox \
    --policy.path=outputs/train/act_duck_pickup_v1/checkpoints/${CHECKPOINT}/pretrained_model