LeRobot YAM Plugins

YAM follower arm + GELLO leader plugins for LeRobot.

Quickstart

Requirements: CAN interface (e.g. can0) and a GELLO leader (port configurable). GELLO assembly information from gello_mechanical.

uv sync
uv run lerobot-teleoperate \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --robot.gripper_type=crank_4310 \
  --teleop.type=yam_leader \
  --teleop.port=/dev/ttyUSB0

Using an existing leader calibration file (for example yam_lerobot.json in the repo root):

uv run lerobot-teleoperate \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --robot.gripper_type=crank_4310 \
  --teleop.type=yam_leader \
  --teleop.port=/dev/ttyUSB0 \
  --teleop.calibration_dir=./ \
  --teleop.id=yam_lerobot

--teleop.id is required when loading an existing calibration file, because LeRobot resolves calibration as <teleop.calibration_dir>/<teleop.id>.json.

GELLO hardware profiles and active assistance

yam_leader defaults to the versioned trlc_dk1_v1 hardware profile. It contains the September 2026 system-identification result for the bimanual TRLC-DK1 leaders: encoder signs/ranges/offsets, effective link masses, assisted joints, gravity gain/current limit, and the spring-open trigger. Gravity assistance and gripper return are therefore on out of the box:

uv run lerobot-teleoperate \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --teleop.type=yam_leader \
  --teleop.port=/dev/ttyUSB0 \
  --teleop.id=yam_gello_left

The profile is defined in lerobot_teleoperator_yam_gello/gravity_profiles.py, next to the controller that consumes it—not hidden in an operator command or calibration cache. The bundled gello_software/FACTR URDF supplies kinematics/COMs; the profile supplies DK1 effective dynamics and encoder mapping. These effective masses are fitted model constants, not literal weigh-scale claims, so they intentionally remain in the profile rather than being written into the upstream active-GELLO URDF.

For a different GELLO build, add a new named profile (and optionally a new packaged URDF/MJCF asset) once its sysid is stable. During bring-up, every field can still be overridden through LeRobot config arguments. Select --teleop.gravity_profile=passive for position sensing with all torque disabled.

Safety measures:

• per-motor current is hard-clipped (default 250 mA, configurable up to 500 mA);
• the squeeze trigger uses current-based position mode and pushes toward its physical-open endpoint (80 mA in the DK1 profile), so the operator must squeeze and hold it closed;
• a 200 ms DYNAMIXEL bus watchdog stops output if updates cease;
• assistance latches off on a motor hardware error, a control exception, or temperature reaching 50 °C;
• disconnect writes zero current and disables torque.

Measuring corrections instead of guessing: gello_gravity_sysid

When assist feels wrong (holds at one pose, drives at others), measure the joint's real gravity curve on the bench and fit the corrections:

python -m lerobot_teleoperator_yam_gello.gello_gravity_sysid \
  --port /dev/ttyUSB0 --id yam_gello_left \
  --joint elbow_flex --poses 5 --enable-torque-output

For each operator-arranged pose it locks the other joints (current-capped position hold), bisects the two breakaway edges of the test joint's holding-current interval, and after >= 3 poses fits measured vs model balance currents to print a suggested gravity_joint_offsets_rad delta (phase gap), a distal mass scale (amplitude ratio), and the measured stiction band. Every probe is dumped to a JSON diagnostics file. Spread the test joint's angles as widely as safely possible — clustered poses make the fit ill-conditioned (the tool warns). The test joint moves during measurement; keep the workspace clear, and power-cycle the leader if the process is ever killed ungracefully.

The DK1 profile's signs and offsets are applied in radian space (q_urdf = sign * q_yam + offset) and were identified together with its ranges and effective masses. Do not mix constants across profiles. Use gravity_assist_dry_run=true while identifying a new build: it logs modeled positions, torques, and currents while applying no torque.

Note: this repo expects lerobot 0.4.3 features (plugin discovery in the standard CLIs). If 0.4.3 is not on PyPI, install it from source instead.

Record

uv run lerobot-record \
  --dataset.repo_id=YOUR_USERNAME/yam_teleop \
  --dataset.single_task="Teleop YAM arm (pick/place)" \
  --dataset.fps=30 \
  --dataset.num_episodes=5 \
  --dataset.episode_time_s=60 \
  --dataset.reset_time_s=15 \
  --dataset.push_to_hub=true \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --robot.gripper_type=crank_4310 \
  --teleop.type=yam_leader \
  --teleop.port=/dev/ttyUSB0

Policy Control

uv run lerobot-control \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --policy.path=YOUR_USERNAME/yam-policy

Extras

List cameras:

uv run lerobot-find-cameras opencv

Teleop with cameras + Rerun:

uv run lerobot-teleoperate \
  --robot.type=yam_follower \
  --robot.port=can0 \
  --robot.gripper_type=crank_4310 \
  --robot.cameras='{
    "wrist": {"type": "opencv", "index_or_path": "/dev/video6", "width": 640, "height": 480, "fps": 30},
    "front": {"type": "opencv", "index_or_path": "/dev/video4", "width": 640, "height": 480, "fps": 30}
  }' \
  --teleop.type=yam_leader \
  --teleop.port=/dev/ttyUSB0 \
  --display_data=true

Safety step limits (normalized per‑cycle steps: joints in [-100, 100], gripper in [0, 100]):

--robot.lerobot_max_step=1 \
--robot.lerobot_gripper_max_step=1

Packages

Install packages separately if you only need one component (e.g., follower‑only for policy inference).