ZefanW's picture
Link the JustRUA paper, project page and code
2629247 verified
|
Raw History Blame Contribute Delete
6.39 kB
metadata
pretty_name: RoboDojo Devset Replays
tags:
  - robotics
  - evaluation
configs:
  - config_name: default
    data_files:
      - split: dev
        path: data/devset.jsonl
  - config_name: video_action_pairs
    data_files:
      - split: dev
        path: index/video-actions.jsonl

RoboDojo Devset Replays

Replay data for the 14 methods in the main comparison. Each method uses the same 111 canonical P50_111 cases (42 tasks).

14 × 111 = 1,554 unique method–case records. 4,662 MP4s and 1,554 action files were verified at the artifact revision in verification.json.

Open the replay explorer · Canonical cases · All replays · Per-method manifests · Video-to-action index · Verification report

Follow the actions in a video

The head, left-wrist and right-wrist cameras record the same robot run. All three views share one actions.json in their case directory. The video-to-action index has one row per MP4, with its camera name, replay run ID, video path and matching action path. Both files have SHA-256 checksums.

For example, open the Compact readings replay for insert_tubes, standard, seed 0, scene 0: head video · left-wrist video · right-wrist video · matching actions.

Read events in decision order. The control_start and control_end fields give cumulative robot control steps. When an event supplies camera_frames for the selected camera, its frame range is half-open: start inclusive, end exclusive. The index reports how many events have a usable range for that camera. Use the action order and control counts for the remaining events.

Method (per-method manifest) Cases MP4s Actions Score / 100
Official L3 111 333 111 29.6562
Full history 111 333 111 34.6771
Adaptive steps 111 333 111 37.8542
20-step cap 111 333 111 37.9792
Relative targets 111 333 111 33.9062
Semantic ops 111 333 111 34.6250
Visual demo 111 333 111 39.8646
Text trajectory 111 333 111 35.1771
Self-built library 111 333 111 36.9896
Affordances drawn 111 333 111 36.7396
Compact readings 111 333 111 39.8333
Depth, predicted 111 333 111 42.3958
Affordance panel 111 333 111 42.1146
Coordinate solving 111 333 111 50.3125

Data format

Case IDs identify the task, variant, seed and scene. Each replays/<method_id>/<case_id>/ directory contains head.mp4, left_wrist.mp4, right_wrist.mp4 and actions.json. The main manifest lists their paths, SHA-256 checksums and byte sizes. Scores use the original 0–1 scale.

Each event in actions.json includes a nullable rationale and rationale_source. The text is the public tool call’s original note, or its reason for giving up. Events without an attributable action note carry null. System budget messages and retrospective comments are excluded. Recorded rationale is available for 97,226 of 98,729 events across 1,553 replays. Coverage by method.

Actions use robodojo.replay.actions.v1. Events distinguish requested targets, observed execution counts and known measured state. Zero-control attempts are preserved. Missing measurements and unavailable frame alignments stay null. Code bodies, private logs and hidden model reasoning are excluded.

Provenance and limitations

  • The 13 local methods follow the published dashboard P50_111 selections. Full history includes one explicitly approved supplementary result; its evidence label is retained.
  • Official L3 scores come from the original grid. The grid does not give original run IDs. The replays are unique public Astra Console candidates matched by task, variant, actual layout, score and success. The source policy label is RoboDojo_Agent_L4_Inspect_EEF@astra. The replay run ID is separate from score provenance.
  • Official videos are unchanged copies of the Solomonz Astra rollouts, whose dataset card states Apache-2.0. See the upstream Console and per-case credits and fixed URLs. The upstream public records omit full commanded targets and measured joint poses.
  • Official store_tools_in_toolbox / standard / scene 0 has conflicting raw-trace success and grid outcome. The official score is preserved and flagged. Official classify_objects_by_language / standard / scene 3 has three late decisions beyond recorded frames. Their frame links are null.

Check a downloaded copy

Run python validate.py --index-only to check 14 × 111 identities and per-method manifests. Run python validate.py to additionally verify every MP4 and action file against its SHA-256. Source snapshots.

Paper

These replays accompany JustRUA: A Controlled Study of GPT-6 Robot-Use Agents. Project page · Code · Replay explorer