Run your research in RoboGUI

Start the camera, model, RoboGUI and manimux serve processes using the matching runbook. The GUI then owns daily rollout operation. The task command is the instruction sent to the policy; the experiment name, condition and notes describe your research and do not modify model inputs.

Free rollouts

Enter a task. Expand Research details for optional experiment name, condition and notes. Click Prepare free rollout, then Start rollout. Use Pause / Hold when needed and Finish (the selected robot may offer Finish & Home or Finish without homing). There is no scoring step. Recording follows the experiment's recording.enabled. Prepare can connect hardware and move to the configured start pose.

Study rollouts

Choose Prepare study rollout when you want layout/repeat identity or optional post-rollout evaluation. Without a template, any text layout ID and positive repeat number are accepted; a reference image is optional. After Finish, save an evaluation or click Skip evaluation. Skipping leaves the episode unlabeled.

A template adds only the constraints you select in the experiment YAML:

run:
  experiment_name: bottle-study
  condition: baseline
  notes: ""
  experiment_template:
    name: Three scene variants
    layout_ids: [near, far, cluttered]
    repeats: 5
    require_reference: false

Omit experiment_template for unconstrained research. Omit layout_ids or repeats inside a template to leave that dimension unconstrained. A template constrains study rollouts; free rollouts remain available. It does not change inference, timing, checkpoint, robot limits or the task prompt. Templates do not automatically schedule, randomize or count a study for you.

For the earlier ten-layout study, use string IDs "01" through "10", repeats: 3 and require_reference: true. Existing recorded identities remain readable. Complete configuration example.

If attaching a reference, select it in Reference layout · Top and enable Attach selected reference image. Its ID must match the study layout, or can supply the ID when the optional layout field is empty. Templates can require an image. Existing numeric filenames and named images such as near.png are supported. The reference capture tool accepts a user-entered ID rather than ten fixed slots.

Prepare freezes the identity and selected image fingerprint for this rollout. Changes for the next rollout do not rewrite previous records. Templates come from the loaded experiment; change the YAML and restart the runtime service to change them.

Where records go

Keep generated records under the repository's top-level data/ directory, outside the Python package. From the repository root, an example output is:

data/yam-pi05-rtc/
└── session-.../
    ├── session-manifest.json
    └── rollout-.../
        ├── meta.json
        ├── result.json
        ├── data.zarr/
        ├── events.jsonl
        ├── videos/
        └── evaluation/       # Only when evaluation is saved

run.output_dir selects the parent of session directories. A workstation can use paths.output_dir in its private station file to override it. Relative experiment output paths resolve from the launch directory; relative station paths resolve from the station file. The GUI displays the actual session and episode paths. The experiment name is metadata, not a filesystem path. Records are ignored by Git. Existing experiments retain their explicitly configured output locations.

Review in the GUI

Open Recorded rollouts. The session directory follows a newly connected runtime; you can replace it with an older session's absolute path. Click Refresh episodes, select a finalized episode and inspect its identity, termination reason and label status. This reads files on the machine running RoboGUI, not on the browser's machine.

Choose a replay source:

Source	Meaning
Measured state	Robot feedback recorded on each runtime tick
Scheduled reference	Timeline targets before the executor
Executor command	Commands produced by the executor; not a hardware acknowledgement

Open offline replay creates a separate paused view, normally on the next port (e.g. 8087 beside 8086). Use Play, Frame and Speed. It follows recorded tick times; it does not interpolate or rerun the executor. Choosing another replay replaces that replay view. Live cameras and controls stay in the original GUI. For a remote workstation, expose/tunnel the replay port as well and use that host in the replay URL. Closing RoboGUI also stops its replay view.

This entry point replays trajectories from finalized data.zarr records, including older compatible episodes. It does not replay their videos or raw model outputs. For an independent NPZ trajectory use action replay.

Optional evaluation

Human labels and PRM are available when they answer your research question. result.json.success describes runtime completion, not whether the robot accomplished the task. A missing human label means unreviewed. See record meanings before comparing metrics.