lerobot-robot-ugo-pro 1.1.0

LeRobot follower implementation for the ugo Pro humanoid robot

LeRobot + ugo Pro Integration

This repository packages a full follower implementation that bridges the ugo Pro R&D model with Hugging Face’s LeRobot ecosystem. It complies with the Bring Your Own Hardware (BYOH) conventions so the robot is autodiscovered by lerobot-teleoperate, lerobot-record, etc.

Key pieces live under lerobot_robot_ugo_pro/:

• config_ugo_pro.py — UgoProConfig, a BYOH-registered RobotConfig.
• ugo_pro.py — the Robot subclass with lifecycle, observation/action schemas, and safety hooks.
• telemetry/ — streaming CSV parser (TelemetryParser) and TelemetryFrame dataclass.
• transport/ — UDP telemetry listener and command client with a 10 ms rate limiter.
• follower/mapper.py — leader→joint mapping with follower gain, mirror, and role support.

Design references, UDP specs, and the end-to-end task plan are documented in docs/.

---

Installation

pip install lerobot-robot-ugo-pro
# or while developing:
pip install -e ".[dev]"

The package exposes the ugo_pro robot type via the lerobot.robots entry-point, so once installed it is automatically available to the LeRobot CLI tools. When running straight from a git checkout (before packaging), set PYTHONPATH=$PWD or LEROBOT_THIRD_PARTY_MODULES=lerobot_robot_ugo_pro.ugo_pro so the follower class is importable.

---

Quick Start

Setup Variables

export REPO_ID="username/my_record_name"
export TASK_NAME="Grab the green cube"

Camera Configuration

lerobot-find-cameras

Teleoperation via CLI

lerobot-teleoperate \
  --robot.type=ugo_pro \
  --robot.id=my_ugo_pro \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 15}, left: {type: opencv, index_or_path: 2, width: 1920, height: 1080, fps: 15}, right: {type: opencv, index_or_path: 1, width: 1920, height: 1080, fps: 15}}" \
  --teleop.type=ugo_bilcon \
  --teleop.id=my_ugo_bilcon \
  --display_data=true

Recording

lerobot-record \
  --robot.type=ugo_pro \
  --robot.id=my_ugo_pro \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 15}, left: {type: opencv, index_or_path: 2, width: 1920, height: 1080, fps: 15}, right: {type: opencv, index_or_path: 1, width: 1920, height: 1080, fps: 15}}" \
  --teleop.type=ugo_bilcon \
  --teleop.id=my_ugo_bilcon \
  --display_data=false \
  --dataset.fps=15 \
  --dataset.push_to_hub=false \
  --dataset.repo_id=$REPO_ID \
  --dataset.single_task=$TASK_NAME
  --dataset.num_episodes=2 \
  --resume=true

Dataset Viewer (Local)

lerobot-dataset-viz \
  --repo-id $REPO_ID \
  --episode-index 0

Replay

lerobot-replay \
  --robot.type=ugo_pro \
  --robot.id=my_ugo_pro \
  --dataset.repo_id=$REPO_ID \
  --dataset.episode=0

Policy Training

lerobot-train \
  --dataset.repo_id=$REPO_ID \
  --policy.repo_id=$REPO_ID_MODEL \
  --policy.type=act  \
  --output_dir=outputs/train/model \
  --job_name=modelname \
  --policy.device=cuda \
  --wandb.enable=true \
  --wandb.project=conditional_act_v3 \
  --policy.push_to_hub=false \
  --steps=40000 \
  --batch_size=4 \
  --save_checkpoint=true \
  --save_freq=10000

---

Configuration Highlights

UgoProConfig exposes the knobs needed to match a particular MCU deployment:

Field	Description
telemetry_host / telemetry_port	UDP bind for incoming MCU telemetry (8886 by default).
mcu_host / command_port	MCU IP + port for outbound commands (192.168.4.40:8888).
left_arm_ids / right_arm_ids	Servo ID ordering (defaults to [21-28] for left and [11-18] for right).
follower_role	"dual", "left-only", or "right-only" to restrict which arm receives targets.
mirror_mode	Swap and sign-flip leader actions to mirror across arms.
follower_gain	Blend factor (0→hold current pose, 1→direct leader tracking, intermediate blends).
action_map	Map arbitrary leader keys (e.g. leader.elbow) to servo IDs.
timeout_sec	Telemetry timeout; exceeded → automatic re-send of the last targets for fail-safe.
cameras	Optional CameraConfig dict; validated for width/height/fps per BYOH rules.

Joint limits, default velocity/torque ceilings, and command history depth are also configurable. See docs/lerobot_ugo_pro_design.md for the complete field reference and validation rules.

---

Telemetry & Command Path

• Telemetry: UgoTelemetryClient binds to telemetry_port, decodes CSV packets via TelemetryParser, and stores the latest TelemetryFrame inside a thread-safe JointStateBuffer.
• Observation: UgoPro.get_observation() translates the frame into per-joint keys (joint_<id>.pos_deg, optional vel_raw, cur_raw, target_deg), VSD metadata, status fields, and camera outputs.
• Action: UgoFollowerMapper normalizes leader input into ordered joint targets (applying mirror mode, follower role, and gain), while UgoCommandClient builds the MCU CSV as a single comma-separated target row (no cmd/id/tar/spd/trq/sync headers) and rate-limits to the configured 10 ms cycle.
• Fail-safe: If telemetry stalls beyond timeout_sec, the follower re-sends the last safe targets so the MCU keeps the current posture.

Refer to docs/ugo_arm_monitoring_spec.md for the CSV schema, packet cadence, and vsd metadata semantics.

---

Development & Testing

git clone https://github.com/ugo-plus/lerobot-robot-ugo-pro.git
cd lerobot-robot-ugo-pro
pip install -e ".[dev]"
pytest

Useful commands:

• pytest — runs the targeted suites (parser, mapper, config, UDP client, robot tests with dummy transports).
• mypy lerobot_robot_ugo_pro tests — type-check runtime + test utilities.
• ruff check . / ruff format . — linting/formatting.
• pre-commit install && pre-commit run --all-files — keep local checks aligned with CI (.github/workflows/ci.yml runs ruff + mypy + pytest).

Need reference data while developing? See:

• docs/custom_robot_integration_with_lerobot.md — BYOH checklist mirrored from upstream docs.
• docs/lerobot_ugo_pro_design.md — architecture decisions and module breakdown.
• docs/ugo_arm_monitoring_spec.md — UDP telemetry / command protocol straight from the MCU team.

---

Support & Future Work

The implementation already covers telemetry parsing, follower mapping, UDP transport, and LeRobot integration. Upcoming work includes feedback channels to teleoperators, dataset tooling, and expanded monitoring hooks. Contributions are welcome — please open an issue or PR if you extend the mapper, transport, or docs for your deployment.