lerobot_rosetta

Rosetta's LeRobot framework adapter: dataset writer, policy runner, and gRPC inference servers.

The naming convention is <framework>_rosetta, rosetta's adapter to a policy framework. This package is the rosetta-to-LeRobot direction: rosetta loads it by name via entry points. The opposite direction (the Robot and Teleoperator plugins that LeRobot auto-discovers) lives in lerobot_robot_rosetta and lerobot_teleoperator_rosetta, whose names are forced by LeRobot's lerobot_robot_*/lerobot_teleoperator_* discovery convention. This package deliberately matches neither prefix, so LeRobot's register_third_party_plugins() never imports it. It depends on lerobot_robot_rosetta (for RosettaConfig, and LeRobot discovers the Robot plugin at policy-run time).

Entry points

Group	Name	Target
rosetta.dataset_writers	lerobot	lerobot_rosetta.dataset_writer:LeRobotDatasetWriter
rosetta.policy_runners	lerobot	lerobot_rosetta.policy_runner:LeRobotPolicyRunner
console_scripts	rosetta_policy_server	lerobot_rosetta.policy_server:main
console_scripts	rosetta_classifier_server	lerobot_rosetta.classifier_server:main

Inference servers

Two gRPC servers speak LeRobot's AsyncInference protocol (the RobotClient connects to either unchanged):

• policy_server.py (rosetta_policy_server): preload and cache wrapper over LeRobot's stock lerobot.async_inference.policy_server. The stock server reloads the checkpoint on every client handshake; this one loads it once and reuses it while the requested (policy_type, pretrained_name_or_path, device) is unchanged.
• classifier_server.py (rosetta_classifier_server): reward-classifier variant with the same preload and cache behavior.

Both accept optional preload flags that load the model before the port is bound, so a socket-level readiness check implies the model is ready:

python -m lerobot_rosetta.policy_server --host=127.0.0.1 --port=8080 \
    --policy-type=act --pretrained-name-or-path=my-org/my-policy --policy-device=cuda

LeRobotPolicyRunner starts the appropriate server with these flags at node configure time (launch_local_server: true), so the first run_policy goal is as fast as any other.

Build

colcon build --packages-select lerobot_rosetta

Documentation

How the LeRobot adapter fits together, and the dataset layout it writes: LeRobot integration and LeRobot data model.

Full Rosetta documentation: https://iblnkn.github.io/rosetta/

License

Apache-2.0