Nodes

Both Rosetta nodes read parameter files (params/) as defaults. A launch file exposes the deployment-specific subset as launch arguments (paths, storage format, log level, lifecycle autostart); everything else is set in the params YAML. Run ros2 launch rosetta <launch_file> --show-args to see the options.

episode_recorder_node

Records contract-specified topics to rosbag2. Launch: episode_recorder_launch.py.

Parameter	Default	Description
contract_path	contracts/so_101.yaml	Path to contract YAML (launch arg)
bag_base_dir	datasets/bags	Directory for rosbag output, relative to the launch cwd (launch arg)
storage_id	mcap	Rosbag format: mcap (recommended) or sqlite3 (launch arg)
default_prompt	""	Task label used when a goal leaves prompt empty
default_max_duration_s	0.0	Max episode duration. 0.0 records until stopped
feedback_rate_hz	2.0	Recording feedback publish rate
record_all	true	Record every topic on the graph, not only contract topics
exclude_topics	[]	Regex list of topics to skip when record_all is on
include_topics	[]	Regex list to always record, overriding exclude_topics
embed_contract	true	Embed the contract text into bag metadata

Actions and services: record_episode (action), ~/start_recording, ~/cancel_recording, ~/delete_last_bag.

Topic recording

By default the recorder records every topic on the ROS 2 graph, not just those declared in the contract, so you never lose data you might need later. This behaves like ros2 bag record -a. Contract topics are required to be present. Only /rosout and /parameter_events are excluded automatically; exclude_topics excludes more, and record_all: false records only contract-declared topics.

Cameras are the exception: per camera the recorder keeps one image_transport stream, preferring /compressed > /zstd > /theora > /compressedDepth > raw. image_transport republishers encode only while subscribed, so recording all of them makes the camera node encode every frame several ways at once.

episode_keyboard_node

Keyboard control for the recorder. Launch: episode_keyboard_launch.py, with recorder_ns (default /episode_recorder) and default_prompt arguments.

Key	Action
r / →	Start recording
s / ←	Stop and save
d / ⌫	Discard episode (stop + delete bag)
t	Edit task prompt for the next episode
h / ?	Help
q	Quit

policy_runner_node

Wraps a policy framework’s inference pipeline in ROS 2 actions. Launch: policy_runner_launch.py. The first block is declared by the node; the second by the resolved framework adapter, lerobot_rosetta here.

Parameter	Default	Description
contract_path	contracts/so_101.yaml	Optional when the checkpoint’s dataset embeds one (launch arg)
framework	lerobot	Policy framework adapter, resolved by entry-point name
is_classifier	false	Publish the reward section as the action output
default_prompt	""	Task used when a goal leaves prompt empty
default_max_duration_s	0.0	Max run duration. 0.0 runs until stopped
feedback_rate_hz	2.0	Execution feedback publish rate

LeRobot adapter parameter	Default	Description
pretrained_name_or_path	(see params file)	HuggingFace model ID or local path (launch arg)
server_address	127.0.0.1:8080	Policy server address (launch arg)
policy_type	act	act, smolvla, diffusion, pi0, pi05, etc. (launch arg)
policy_device	cuda	cuda, xpu, mps, cpu, or cuda:0 (falls back to cpu if unavailable)
actions_per_chunk	30	Actions per inference chunk
chunk_size_threshold	0.95	When to request a new chunk (0.0-1.0)
aggregate_fn_name	weighted_average	weighted_average, latest_only, average, conservative
launch_local_server	true	Auto-start the policy server at configure, with model preload (launch arg)
server_startup_timeout_sec	120.0	Max wait for the server, covering model preload
obs_similarity_atol	-1.0	Observation filtering tolerance. Ignored by stock LeRobot v0.6.0, where the filter is hardcoded on

With use_sim_time true the adapter paces observations and actions on the sim clock at the contract fps.

Actions and services: run_policy (action), ~/start_policy, ~/cancel_policy. A launch namespace prefixes the action name, as in /robot_policy/run_policy under hil_launch.py.

When contract_path is empty, the node resolves the contract from the checkpoint: pretrained_name_or_path → train_config.json → the training dataset → meta/rosetta_contract.yaml.

hil_manager_node

Orchestrates human-in-the-loop episodes. Launch: hil_launch.py, which wires the manager, a policy runner (namespace robot_policy), an optional reward classifier (namespace reward_classifier), and the recorder. Params: params/hil_manager.yaml covers all four.

Action: manage_episode. Services: ~/start_episode, ~/end_episode and ~/set_intervention and ~/set_reward_override (SetBool), ~/cancel_episode and ~/clear_reward_override (Trigger).

~/end_episode is the deliberate, labelled end: the goal succeeds and the verdict lands in outcome. ~/cancel_episode abandons the take. Neither deletes a bag.

ROS 2 lifecycle

All Rosetta nodes are lifecycle nodes.

Transition	Effect
configure	Create subscriptions (start buffering), create publishers (disabled)
activate	Enable publishers, start watchdog, open goal acceptance
deactivate → cleanup	Safety action, disable publishers, destroy resources

Goals are accepted only while active, one at a time per node. A deactivate stops in-progress work and ends its goal ABORTED with termination_reason: node_deactivated.