SO-ARM101 Bi-Arm (Dual Follower) Tutorial

Buy in Store

This guide introduces the complete workflow for training a bi-arm SO-ARM robot system with LeRobot, including hardware connection, bi-arm calibration, bi-arm teleoperation, dataset recording and management, ACT policy training, and real-robot deployment. Following this guide, you can use two leader arms and two follower arms to collect demonstration data, train an imitation-learning policy, and run it on real robotic arms.

First, connect the cables as follows:

Role	Port
Left follower arm	/dev/ttyACM0
Right follower arm	/dev/ttyACM1
Left leader arm	/dev/ttyACM2
Right leader arm	/dev/ttyACM3

The follower-arm type is so101_follower, and the leader-arm type is so101_leader (in LeRobot, so100_leader and so101_leader share the same implementation).

Prerequisites

Install Dependencies

For environment setup, see the SO-ARM101 Tutorial.

USB Permissions

sudo chmod 666 /dev/ttyACM0 /dev/ttyACM1 /dev/ttyACM2 /dev/ttyACM3

1. Calibration (Critical Step)

1.1 Calibrate the Left Follower Arm

lerobot-calibrate \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.id=my_so101_bi_follower_left

1.2 Calibrate the Right Follower Arm

lerobot-calibrate \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower_right

1.3 Calibrate the Left Leader Arm

lerobot-calibrate \
  --teleop.type=so101_leader \
  --teleop.port=/dev/ttyACM2 \
  --teleop.id=my_so101_bi_leader_left

1.4 Calibrate the Right Leader Arm

lerobot-calibrate \
  --teleop.type=so101_leader \
  --teleop.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader_right

After calibration, the files are saved to:

~/.cache/huggingface/lerobot/calibration/robots/so_follower/my_so101_bi_follower_left.json
~/.cache/huggingface/lerobot/calibration/robots/so_follower/my_so101_bi_follower_right.json
~/.cache/huggingface/lerobot/calibration/teleoperators/so_leader/my_so101_bi_leader_left.json
~/.cache/huggingface/lerobot/calibration/teleoperators/so_leader/my_so101_bi_leader_right.json

Directory naming: so101_follower and so100_follower, so101_leader and so100_leader share the same implementation, so the directories are uniformly so_follower / so_leader; the leader arm is a teleoperator, so its calibration files go under teleoperators/ rather than robots/.

(Optional) If You Previously Calibrated with Other IDs

For example, if you previously used my_awesome_follower_arm1, my_awesome_follower_arm2, etc., you can copy the calibration files:

CAL_DIR=~/.cache/huggingface/lerobot/calibration

cp $CAL_DIR/robots/so_follower/my_awesome_follower_arm1.json \
   $CAL_DIR/robots/so_follower/my_so101_bi_follower_left.json

cp $CAL_DIR/robots/so_follower/my_awesome_follower_arm2.json \
   $CAL_DIR/robots/so_follower/my_so101_bi_follower_right.json

cp $CAL_DIR/teleoperators/so_leader/my_awesome_leader_arm3.json \
   $CAL_DIR/teleoperators/so_leader/my_so101_bi_leader_left.json

cp $CAL_DIR/teleoperators/so_leader/my_awesome_leader_arm4.json \
   $CAL_DIR/teleoperators/so_leader/my_so101_bi_leader_right.json

2. Bi-Arm Teleoperation

2.1 Without Cameras

lerobot-teleoperate \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM2 \
  --teleop.right_arm_config.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader \
  --display_data=true

2.2 With Cameras

Use lerobot-find-cameras opencv to list camera indices; you can also add or remove cameras as needed.

lerobot-teleoperate \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM2 \
  --teleop.right_arm_config.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader \
  --display_data=true

Safety Note

• Be aware of the surroundings and avoid follower-arm collisions.

3. Record a Dataset

3.1 Save Locally (No Hub Upload)

Add --dataset.root (data goes to that directory) and --dataset.push_to_hub=false, plus --dataset.no_stamp=true to keep the dataset name stable (otherwise a timestamp is automatically appended to repo_id, and later resume recording / replay / training will not find it).

Note: repo_id should contain a / (in the form username/dataset_name); local datasets are not actually uploaded.

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM2 \
  --teleop.right_arm_config.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader \
  --dataset.repo_id=juxi/bi_so101_task \
  --dataset.root=./datasets/bi_so101_task \
  --dataset.push_to_hub=false \
  --dataset.no_stamp=true \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.num_episodes=50 \
  --dataset.fps=30 \
  --dataset.episode_time_s=30 \
  --dataset.reset_time_s=10 \
  --dataset.video=true \
  --display_data=true

Video encoding already defaults to libsvtav1, so there is no need to specify it; to customize, use nested parameters like --dataset.rgb_encoder.vcodec=h264.

The data is saved under ./datasets/bi_so101_task/, with this structure:

├── meta/
│   ├── info.json         # dataset info (fps, feature shapes, etc.)
│   ├── episodes/         # per-episode metadata (chunk-000/...)
│   ├── stats.json        # per-feature normalization statistics
│   └── tasks.parquet     # task text → task_index
├── data/                 # per-frame feature data (chunk-*.parquet)
└── videos/               # one subdirectory per camera (chunk-*.mp4)

3.2 Upload to Hugging Face Hub

If you want automatic upload, keep HF_USER and drop root and push_to_hub=false (upload is the default). Make sure the ports and camera indices match the wiring table:

export HF_USER=your_hf_username

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM2 \
  --teleop.right_arm_config.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader \
  --dataset.repo_id=${HF_USER}/bi_so101_task \
  --dataset.no_stamp=true \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.num_episodes=50 \
  --dataset.fps=30 \
  --dataset.episode_time_s=30 \
  --dataset.reset_time_s=10 \
  --dataset.video=true \
  --display_data=true

The uploaded Hub repository name will be ${HF_USER}/bi_so101_task, matching the repo_id used for training from the Hub in 4.2 below. A local copy is first saved to ~/.cache/huggingface/lerobot/${HF_USER}/bi_so101_task/.

3.3 Continue Recording (Resume)

If recording exits unexpectedly (for example, you quit with the right-click while in the reset phase), or you want to complete the collection in several sessions, you can use --resume to keep appending episodes to the same dataset.

Note:

• You must add --resume=true, otherwise LeRobotDataset.create() errors out because the directory already exists.
• The resume command's --dataset.root and --dataset.repo_id must exactly match the first recording (3.1) (resume requires an explicit root).
• --dataset.num_episodes is how many episodes to record this time, not the total goal. For example, if you have recorded 15 and want 50 in total, write 35.
• When quitting, try to do so during episode recording or right after it ends naturally; avoid quitting during the "Reset the environment" phase (it makes saving an empty episode fail).

lerobot-record \
  --resume=true \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM2 \
  --teleop.right_arm_config.port=/dev/ttyACM3 \
  --teleop.id=my_so101_bi_leader \
  --dataset.repo_id=juxi/bi_so101_task \
  --dataset.root=./datasets/bi_so101_task \
  --dataset.push_to_hub=false \
  --dataset.no_stamp=true \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.num_episodes=35 \
  --dataset.fps=30 \
  --dataset.episode_time_s=30 \
  --dataset.reset_time_s=10 \
  --dataset.video=true \
  --display_data=true

3.4 Replay and Delete Episodes

Replay a Specific Episode

lerobot-replay \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --dataset.repo_id=juxi/bi_so101_task \
  --dataset.root=./datasets/bi_so101_task \
  --dataset.episode=24

episode is a 0-based index; 24 means the 25th episode.

Delete a Specific Episode

python -m lerobot.scripts.lerobot_edit_dataset \
  --repo_id=juxi/bi_so101_task \
  --root=./datasets/bi_so101_task \
  --operation.type=delete_episodes \
  --operation.episode_indices="[24]"

The deletion rewrites the dataset in place, and the original data is backed up to ./datasets/bi_so101_task_old/. Once you have confirmed the new dataset is correct, you can delete the backup manually:

rm -rf ./datasets/bi_so101_task_old

Delete the Entire Dataset

rm -rf ./datasets/bi_so101_task

4. ACT Training

4.1 Train from a Local Dataset

lerobot-train \
  --dataset.repo_id=juxi/bi_so101_task \
  --dataset.root=./datasets/bi_so101_task \
  --policy.type=act \
  --policy.device=cuda \
  --steps=60000 \
  --output_dir=outputs/train/act_bi_so101 \
  --wandb.enable=false \
  --policy.push_to_hub=false

--dataset.root points to the dataset directory recorded in 3.1 (repo_id must match the recording). If the --output_dir directory already exists, it raises FileExistsError; use a new output directory or add --resume=true to continue training.

4.2 Train from Hugging Face Hub

export HF_USER=your_hf_username

lerobot-train \
  --dataset.repo_id=${HF_USER}/bi_so101_task \
  --policy.type=act \
  --policy.device=cuda \
  --steps=100000 \
  --output_dir=outputs/train/act_bi_so101 \
  --wandb.enable=false \
  --policy.push_to_hub=false

The command above uses ACT's default parameters (chunk_size=100, dim_model=512, etc.).

repo_id must match the repository name used in the 3.2 upload (3.2 already added --dataset.no_stamp=true, so the repository name is fixed as ${HF_USER}/bi_so101_task). No --dataset.root is needed for training — it downloads from the Hub automatically.

5. Real-Robot Deployment

Note: lerobot-record is only for collecting demonstration data. To deploy a trained policy, use lerobot-rollout — the current version of lerobot-record no longer accepts --policy.path and also rejects dataset names with the eval_ prefix.

5.1 Live Evaluation (No Data Recording)

lerobot-rollout \
  --strategy.type=base \
  --policy.path=outputs/train/act_bi_so101/checkpoints/last/pretrained_model \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --task="Pick the cube with left arm and hand it to right arm" \
  --duration=60 \
  --display_data=true

• --duration is the run time in seconds; 0 means no time limit.
• To take over/stop mid-run, add --interactive=true and control it in the terminal with commands like /stop and /reset.

5.2 Evaluate and Record Data (Local)

Use the episodic strategy (behavior similar to the old lerobot-record: records episode by episode with a reset phase):

lerobot-rollout \
  --strategy.type=episodic \
  --policy.path=outputs/train/act_bi_so101/checkpoints/last/pretrained_model \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --dataset.repo_id=juxi/rollout_bi_so101_task \
  --dataset.root=./datasets/rollout_bi_so101_task \
  --dataset.no_stamp=true \
  --dataset.num_episodes=10 \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.fps=30 \
  --display_data=true

Deployment dataset names must start with rollout_ (an enforced convention in the current version). When recording locally, adding --dataset.root and --dataset.no_stamp=true is recommended to keep a timestamp from being appended to the directory name.

5.3 Upload Evaluation Data to Hugging Face Hub

export HF_USER=your_hf_username

lerobot-rollout \
  --strategy.type=episodic \
  --policy.path=outputs/train/act_bi_so101/checkpoints/last/pretrained_model \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_so101_bi_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30}
  }' \
  --dataset.repo_id=${HF_USER}/rollout_bi_so101_task \
  --dataset.no_stamp=true \
  --dataset.num_episodes=10 \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.fps=30 \
  --display_data=true

6. FAQ

Problem	Cause	Solution
Teleoperation prompts for re-calibration	bi_so_follower cannot find calibration files with the _left / _right suffix	Re-calibrate with IDs containing _left / _right, or copy the existing calibration files
Leader arm cannot be moved	Leader torque not disabled	Re-calibrate or check the motor
Resume recording reports the directory already exists	--resume=true was not added	Add --resume=true to the lerobot-record command
--resume=true errors and demands root	Resume must specify the dataset directory explicitly	Add --dataset.root=./datasets/bi_so101_task to the resume command, matching the first recording
Dataset directory name has an extra timestamp, and replay/training cannot find it	no_stamp was not set during recording, so a timestamp was automatically appended to repo_id	Add --dataset.no_stamp=true when recording/resuming
--dataset.vcodec=... reports the parameter does not exist	Legacy parameter; video encoding parameters are now nested	Use --dataset.rgb_encoder.vcodec=h264 instead (the default is already libsvtav1)
During deployment, lerobot-record reports --policy.path / eval_ errors	The current version of lerobot-record no longer has policy deployment capability	Use lerobot-rollout --strategy.type=episodic for deployment, with dataset names starting with rollout_
Left and right arms are swapped	Wrong port configuration	Swap left_arm_config.port and right_arm_config.port
Training cannot find the dataset	The local dataset's root was not specified	Add --dataset.root=./datasets/xxx when training
The dataset gets uploaded automatically	push_to_hub=false was not set	Add --dataset.push_to_hub=false when recording
On exit it reports You must add one or several frames before calling add_episode	Exited during the reset phase; the current episode has no frames	Does not affect already-recorded data; continue recording with --resume=true

Related Products

SO-ARM101 Developer Kit→