OneWorld-AI/Real2Gym_demo_videos

Real2Gym demo videos

Full-resolution 3x3 comparison renders for the Real2Gym native reconstruction audits. They are held here instead of in the report Space, which has a 1 GB repository limit.

Each file is one episode. Rows are the head, left wrist and right wrist cameras; columns are the real recording, the native MuJoCo simulation, and the real recording with the simulated mesh boundaries drawn in green. Every file covers its complete source episode at the recorded frame rate, with no success-only trimming and no re-encode.

<task>/training_<NN>_<episode prefix>_3x3.mp4
<task>/validation_<NN>_<episode prefix>_3x3.mp4

manifest.json lists every file with its episode id, split, byte size, sha256, frame count, frame rate and pixel size. The same records travel with the report as <task>/videos_manifest.json.

This repository is public. Comparison videos stream directly inside the report HTML without a Hugging Face token. Source recordings: XDOF/ABC-130k.

RoboSimStudio approach · 15 October 2026

put_the_screwdriver_in_the_bin_rss/ contains 15 additional, unmodified comparisons. Their simulation column comes from RoboSimStudio / Newton / MuJoCo Warp, using inherited training geometry and newly simulated trajectories. Earlier approach files remain unchanged. The corresponding five-tab report is in the Real2Gym Space under the same folder name.

RoboSimStudio mixed-dishes approach · 17 October 2026

place_the_mixed_dishes_into_the_plastic_bin_rss/ contains 15 additional, unmodified comparisons. Their simulation column comes from RoboSimStudio / Newton / MuJoCo Warp, using the station and dish shapes inherited from the stopped 15 October mixed-dishes candidate and newly simulated trajectories. Earlier approach files remain unchanged. The corresponding five-tab report is in the Real2Gym Space under the same folder name.

Native RoboSimStudio Lego validation · 19 October 2026

sort_the_legos_into_containers_by_color_rss/ contains two unchanged 3x3 validation comparisons: e930013f and fade8237. The native RoboSimStudio scenes were reconstructed independently using first-frame visual inputs and actions. Both rollouts failed (0/2). Future real frames appear only in the blind final comparison, after both evaluations were sealed. The corresponding report is in the Real2Gym Space.

Frame-0-only held-out evaluations · 2 November 2026

The original 16-task Real2Gym batch has 1 native success and 15 unsuccessful native runs. Three of the unsuccessful runs (phone removal, hats and shoes) remain inconclusive for generalization because initial support failed. Instrument setup is conditional on unresolved accessory/metronome/cover inventory.

Each task uses a validation episode distinct from reconstruction training. Agents used only exported frame 0 and initial simulator renders for visual adaptation. Recorded actions were replayed with native dynamics and the existing task criteria. No later real or simulated frames were viewed, and no rollout videos were produced for this batch. The older comparison videos above are separate evidence.

Materials live under evaluations/2026-11-02_first-frame/:

• Per-task results (CSV) and batch manifest (JSON).
• Protocol and interpretation limits.
• reports/: canonical final reports, with local roots replaced by path placeholders.
• frame0/: unchanged source top-camera first frames.
• archives/: evaluation evidence snapshots, including scripts, model descriptions, numerical traces, initial imagery, prior runs and repair audits.
• files/: per-file and archive SHA256 inventories.

The snapshots retain original file bytes and original workspace paths. They are not self-contained simulator distributions: original scripts may require external training assets and raw ABC streams. The canonical verdict is snapshot/summary.json; nested superseded summaries are historical. Asset file symlinks are materialized; cache/bytecode files are omitted. The top-level manifest.json remains the older video manifest; this evaluation batch has its own manifest linked above.

Evaluation report in ABC_simulators.