Complete Guide to Double-Arm SO-ARM Training

Introduction

This guide walks through the complete pipeline for training a dual-arm SO-ARM robotic system with LeRobot. It covers hardware setup, arm calibration, dual-arm teleoperation, dataset recording and management, ACT policy training, and real-robot deployment. By following these steps, you can collect demonstration data with two leader arms and two follower arms, train an imitation-learning policy, and deploy it on the real robot.

First, connect the wires 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).

---

0. Prerequisites

0.1 Install Dependencies

See related links.

0.2 USB Permissions

sudo chmod 666 /dev/ttyACM1 /dev/ttyACM2 /dev/ttyACM3 /dev/ttyACM4

1. Calibration (Key Step)

1.1 Calibrate the Left Follower Arm

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

1.2 Calibrate the Right Follower Arm

lerobot-calibrate \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM2 \
  --robot.id=my_awesome_bimanual_follower_right

1.3 Calibrate the Left Leader Arm

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

1.4 Calibrate the Right Leader Arm

lerobot-calibrate \
  --teleop.type=so101_leader \
  --teleop.port=/dev/ttyACM4 \
  --teleop.id=my_awesome_bimanual_leader_right

After calibration, the files will be saved at:

~/.cache/huggingface/lerobot/calibration/robots/so101_follower/my_awesome_bimanual_follower_left.json
~/.cache/huggingface/lerobot/calibration/robots/so101_follower/my_awesome_bimanual_follower_right.json
~/.cache/huggingface/lerobot/calibration/robots/so101_leader/my_awesome_bimanual_leader_left.json
~/.cache/huggingface/lerobot/calibration/robots/so101_leader/my_awesome_bimanual_leader_right.json

(Optional) If You Have 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/robots

cp $CAL_DIR/so101_follower/my_awesome_follower_arm1.json \
   $CAL_DIR/so101_follower/my_awesome_bimanual_follower_left.json

cp $CAL_DIR/so101_follower/my_awesome_follower_arm2.json \
   $CAL_DIR/so101_follower/my_awesome_bimanual_follower_right.json

cp $CAL_DIR/so101_leader/my_awesome_leader_arm3.json \
   $CAL_DIR/so101_leader/my_awesome_bimanual_leader_left.json

cp $CAL_DIR/so101_leader/my_awesome_leader_arm4.json \
   $CAL_DIR/so101_leader/my_awesome_bimanual_leader_right.json

---

2. Dual-Arm Teleoperation

2.1 Without Camera

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

2.2 With Camera

lerobot-teleoperate \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_awesome_bimanual_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_awesome_bimanual_leader \
  --display_data=true

You can use lerobot-find-cameras opencv to view camera indexes.

Safety Tips

• Be aware of the surroundings to avoid collisions of the follower arms.

3. Record Dataset

3.1 Save Locally (Do Not Upload to Hub)

Add --dataset.root and --dataset.push_to_hub=false.

Note: repo_id must contain /. For local datasets, you can use local/ as a placeholder prefix; it will not actually upload.

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_awesome_bimanual_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_awesome_bimanual_leader \
  --dataset.repo_id=seeed/bimanual_so101_task \
  --dataset.push_to_hub=false \
  --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 \
  --dataset.vcodec=libsvtav1 \
  --display_data=true

Data will be saved to ~///bimanual_so101_task/, structured as follows:

./datasets/bimanual_so101_task/
├── meta/
│   ├── info.json
│   ├── episodes/
│   ├── stats/
│   └── tasks/
├── data/
└── videos/

3.2 Upload to Hugging Face Hub

If you want to upload automatically, keep HF_USER and remove root and push_to_hub=false:

export HF_USER=your_hf_username

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM1 \
  --robot.right_arm_config.port=/dev/ttyACM2 \
  --robot.id=my_awesome_bimanual_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 1, "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=bi_so_leader \
  --teleop.left_arm_config.port=/dev/ttyACM3 \
  --teleop.right_arm_config.port=/dev/ttyACM4 \
  --teleop.id=my_awesome_bimanual_leader \
  --dataset.repo_id=${HF_USER}/bimanual_so101_task \
  --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 \
  --dataset.vcodec=libsvtav1 \
  --display_data=true

3.3 Continue Recording (Resume)

If the recording exits unexpectedly (for example, pressing the right button to exit during the reset phase), or if you want to complete the collection in multiple sessions, you can use --resume to continue appending episodes to the same dataset.

Notes:

• You must add --resume=true; otherwise LeRobotDataset.create() will report an error because the directory already exists.
• --dataset.num_episodes refers to how many episodes to record this time, not the total target. For example, if 15 episodes have already been recorded and you want to reach 50, set it to 35.
• Try to exit during an episode recording or after a natural end; avoid exiting during the "Reset the environment" phase (which may cause empty episode save failures).

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_awesome_bimanual_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_awesome_bimanual_leader \
  --dataset.repo_id=seeed/bimanual_so101_task \
  --dataset.push_to_hub=false \
  --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 \
  --dataset.vcodec=libsvtav1 \
  --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_awesome_bimanual_follower \
  --dataset.repo_id=seeed/bimanual_so101_task \
  --dataset.episode=24

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

Delete a Specific Episode

python -m lerobot.scripts.lerobot_edit_dataset \
  --repo_id=seeed/bimanual_so101_task \
  --operation.type=delete_episodes \
  --operation.episode_indices="[24]"

After deletion, the dataset will be rewritten in place, and the original data will be backed up to ./datasets/bimanual_so101_task_old/. After confirming the new dataset is correct, you can manually delete the backup:

rm -rf ./datasets/bimanual_so101_task_old

Note: If you are using an older version of LeRobot before the fix, pointing --root to the local dataset root directory may fail due to path parsing errors. The lerobot_edit_dataset.py in the current project has been fixed for this scenario.

Delete the Entire Dataset

rm -rf ./datasets/bimanual_so101_task

---

4. ACT Training

4.1 Train from Local Dataset

lerobot-train \
  --dataset.repo_id=seeed/bimanual_so101_task \
  --policy.type=act \
  --policy.device=cuda \
  --steps=60000 \
  --output_dir=outputs/train/act_bimanual_so101 \
  --wandb.enable=false \
  --policy.push_to_hub=false

4.2 Train from Hugging Face Hub

export HF_USER=your_hf_username

lerobot-train \
  --dataset.repo_id=${HF_USER}/bimanual_so101_task \
  --policy.type=act \
  --policy.device=cuda \
  --steps=100000 \
  --output_dir=outputs/train/act_bimanual_so101 \
  --wandb.enable=false \
  --policy.push_to_hub=false

The above uses ACT's default parameters (chunk_size=100, dim_model=512, etc.). If the dataset is small (for example, fewer than 50 episodes), you can explicitly reduce the model size to lower the risk of overfitting, e.g. --policy.chunk_size=50 --policy.dim_model=256 --batch_size=16 --steps=30000.

---

5. Real-Robot Deployment

5.1 Save Evaluation Data Locally

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM0 \
  --robot.right_arm_config.port=/dev/ttyACM1 \
  --robot.id=my_awesome_bimanual_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=local/eval_bimanual_so101_task8 \
  --dataset.root=./datasets/eval_bimanual_so101_task8 \
  --dataset.push_to_hub=false \
  --dataset.num_episodes=10 \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.fps=30 \
  --dataset.video=true \
  --policy.path=outputs/train/act_bimanual_so101/checkpoints/last/pretrained_model \
  --display_data=true

5.2 Upload to Hugging Face Hub

export HF_USER=your_hf_username

lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_config.port=/dev/ttyACM1 \
  --robot.right_arm_config.port=/dev/ttyACM2 \
  --robot.id=my_awesome_bimanual_follower \
  --robot.left_arm_config.cameras='{
    left_wrist: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}
  }' \
  --robot.right_arm_config.cameras='{
    right_wrist: {"type": "opencv", "index_or_path": 1, "width": 640, "height": 480, "fps": 30}
  }' \
  --dataset.repo_id=${HF_USER}/eval_bimanual_so101_task \
  --dataset.num_episodes=10 \
  --dataset.single_task="Pick the cube with left arm and hand it to right arm" \
  --dataset.fps=30 \
  --dataset.video=true \
  --policy.path=outputs/train/act_bimanual_so101/checkpoints/last/pretrained_model \
  --display_data=true

---

6. FAQ

Problem	Cause	Solution
Teleoperation prompts to recalibrate	bi_so_follower cannot find calibration files with _left / _right suffix	Recalibrate with IDs containing_left / _right, or copy existing calibration files
Leader arm cannot be moved by hand	Leader torque is not disabled	Recalibrate or check the motors
"Directory already exists" when continuing recording	--resume=true was not added	Add--resume=true to the lerobot-record command
Left and right arms are swapped	Port configuration error	Swapleft_arm_config.port and right_arm_config.port
Dataset not found during training	root was not specified for the local dataset	Add--dataset.root=./datasets/xxx during training
Dataset is uploaded automatically	push_to_hub=false was not set	Add--dataset.push_to_hub=false during recording
Exits withYou 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; the current code has fixed this scenario and empty episodes will be automatically skipped