Training a Model

Train and finetune state-of-the-art robot policies

Get the repository

GitHub - anvil-robotics/anvil-embodied-ai: Anvil-Embodied-AI: A robust mission-control platform for Physical AI. Providing the essential infrastructure to deploy, orchestrate, and optimize advanced AI models on Anvil Robotics' robot platforms. Explore more at https://anvil.bot/GitHub

Anvil's embodied AI SDK is in early beta. We will be improving reliability and adding support for new models and workflows soon.

Requirements

• Python 3.12+

• uv

• ffmpeg

Set up environment

uv sync --all-packages

Convert data to train-ready format

Anvil uses the standard LeRobotDataset v3.0 for training, and so the recorded Anvil MCAP must be converted before training. To do this, Anvil provides the mcap_converter package, which will go through your recorded data and compile the episodes into a single LeRobotDataset that's ready for training:

uv run mcap-convert --input-dir data/raw/take-apart-legos --output-dir data/datasets/legos --config configs/mcap_converter/openarm_bimanual_quest.yaml

Train

Anvil also leverages LeRobot for training support, and wraps it for convenience with our lerobot_training package, that makes it easy to use with Anvil data.

uv run lerobot-train --dataset.root=data/datasets/legos --policy.type=act --dataset.repo_id=local --policy.push_to_hub=false