Configure a deployment

Start from a complete experiment YAML. It composes existing components; you do not need to write Python to select another supported checkpoint, camera set or executor.

What you are changing	Owner
Task prompt, output directory, rollout budget, optional study template	Experiment run
Robot assembly and command frequency	Experiment robot
Device addresses, serials and local checkpoint paths	Private station
Geometry, joint order, tool mounting and TCP	Embodiment component / assembly YAML
Camera resolution, FPS and optional exposure overrides	Sensor / camera-set YAML
Model input camera mapping and action conversion	Experiment policy.adapter
Checkpoint, model configuration and normalization	Policy recipe
Model action spacing and predicted horizon	Experiment policy
Request cadence and chunk handoff	inference
Command generation, smoothing and configured limits	executor / shared control profile
Scene layout and display-only initial poses	RoboGUI YAML

See the annotated experiment and the configuration field reference.

References and overrides

config: references resolve relative to the YAML containing them. Mappings merge; lists replace rather than concatenate. The private station binds installation values without changing policy action semantics or execution settings. Its local paths resolve relative to the station file. Relative experiment output paths resolve from the directory where you launch the process.

Use --local .local/station_lab.yaml consistently across processes that support station loading. Otherwise startup selects the experiment's local: reference, then manimux/configs/local/station.yaml. See the entry-point scope table.

Keep the clocks separate

policy.action_dt_s is the spacing between model trajectory points. robot.control_hz is the command-loop frequency. Camera FPS is a third setting. A 30 Hz model trajectory can be linearly interpolated by Timeline and sampled by a 100 Hz command loop. Neither setting guarantees measured physical tracking speed.

policy.horizon_policy_steps is the model output length. inference.chunk_policy_steps controls the selected scheduling method's execution window or query cadence; its exact meaning is method dependent. Diffusion sampling steps are a separate model-side setting.

Apply changes

Edit the owning YAML and restart the affected service. Editing a file does not update a process already using its resolved configuration. RoboGUI changes only the controls it explicitly exposes; it does not hot-reload all configuration.

For a new component or behavior, follow the development integration map.