lerobot-teleopit

Convert Teleopit teleoperation recordings into LeRobot Datasets, train ACT or GR00T N1.7 policies, and run a policy service on a host workstation.

Teleopit recording -> LeRobot Dataset -> ACT / GR00T -> Policy Server -> Teleopit Onboard

Real-time robot control remains the responsibility of Teleopit. This repository handles data conversion, model training, and action-plan generation on the host; it does not control the robot motors directly.

Installation

uv is required. LeRobot 0.6.0 requires Python 3.12 or later.

git clone https://github.com/BotRunner64/lerobot-teleopit.git
cd lerobot-teleopit

Install the dependency groups needed for your use case:

# Convert datasets or run ReplayPolicy
uv sync --extra dataset

# Train or deploy ACT
uv sync --extra dataset --extra train

# Train or deploy GR00T N1.7
uv sync --extra dataset --extra train --extra groot

Then activate the project environment:

source .venv/bin/activate

For development, append --extra dev to the appropriate command above. The project currently pins LeRobot to version 0.6.0.

Quick start

Convert a Teleopit recording into a LeRobot Dataset:

python scripts/convert_dataset.py \
  --source data/raw/my_task \
  --output data/lerobot/my_task \
  --repo-id local/my_task

Train ACT on the first GPU:

python scripts/train_policy.py \
  --policy act \
  --dataset-root data/lerobot/my_task \
  --devices 0

After training, connect the onboard runtime with ReplayPolicy first, then switch to the generated checkpoint. See the documents below for complete commands and options.

Documentation

Topic	English	中文
Dataset conversion and training	English	中文
Policy deployment	English	中文
Action space and root-pose transforms	English	中文

Each document pair contains the same technical content in both languages.

Development checks

uv run ruff check .
uv run pytest -q