Rosetta

Rosetta connects your ROS 2 robot to robot-learning frameworks such as LeRobot.

Your robot publishes data streams. A policy wants frames. The contract is a YAML file that defines this transformation from streams to frames. Rosetta applies the transforms defined in the contract identically for both data preparation and when you go to run your policy live.

robot_type: so_arm101
robot_interface: ros2
fps: 50
observations:
  observation.images.wrist:                          # key: what the model calls it
    channel: {topic: /wrist_camera/image_raw,        # channel: topic, type, QoS
              type: sensor_msgs/msg/Image}
    align: {strategy: hold, timeline: header}        # align: when
    apply: [resize: [480, 480]]                      # apply: how values change
  observation.state:
    channel: {topic: /joint_states, type: sensor_msgs/msg/JointState}
    align: {strategy: hold, timeline: header}
    select: [position.shoulder_pan_joint, position.elbow_flex_joint]   # select: which fields
    apply: [rad2deg]
actions:
  action:
    channel: {topic: /forward_position_controller/commands,
              type: std_msgs/msg/Float64MultiArray, safety: hold}
    align: {strategy: hold, timeline: receive}
    select: [shoulder_pan.pos, elbow_flex.pos]
    apply: [clamp: {min: -3.14159, max: 3.14159}, rad2deg]

Five steps take you from data collection to running a policy on your robot.

ros2 launch rosetta_ros episode_recorder_launch.py contract_path:=robot.yaml                 # 1 record
$EDITOR robot.yaml                                                                           # 2 define
ros2 run rosetta_ros rosetta_port --raw-dir datasets/bags --contract robot.yaml --repo-id my # 3 prepare
lerobot-train --dataset.repo_id=my --policy.type=act --policy.push_to_hub=false \
    --output_dir=outputs/train/my                                                            # 4 train
ros2 launch rosetta_ros policy_runner_launch.py contract_path:=robot.yaml \
    pretrained_name_or_path:=outputs/train/my/checkpoints/last/pretrained_model              # 5 deploy

When you change the contract you redo steps 2 to 5. The bags from step 1 stay.

New here? Install, then follow From bags to a moving arm.

Getting started installs Rosetta and walks the whole workflow once, on a simulated arm.

Getting started

• Install
• From bags to a moving arm

How-to guides cover one task each: recording, writing a contract, preparing a dataset, training, deploying, routing commands through a mux, human-in-the-loop episodes, and extending Rosetta with a message type, a framework or a robot interface.

How-to guides

• Record episodes
• Write a contract
• Prepare a dataset
• Train a policy
• Deploy a policy
• Route commands through a mux
• Run human-in-the-loop episodes
• Add a message type
• Add a framework
• Add a robot interface

Reference lists the exact contract keys, message types, CLI flags, node interfaces and Python API.

Reference

• Glossary
• Contract
• Message types and operators
• rosetta_port
• Nodes
• Interfaces
• Python API
• Packages

Explanation covers why Rosetta records raw and keeps the translation in one file.

Explanation

• About the contract