DropperLab
Auditable dual-arm pipette simulation artifacts. Public research archive, not a physics-qualified benchmark result.
The delivered baseline is a 75.701475-second accepted-state trajectory: approximately 1 mL aspirated, 1.008072 mL retained in the tube, both tools racked and released, and both arms home. The 75.83-second video samples real recorded states at 12 fps.
Provenance
Thirteen frozen cases include explicit numerical/controller repairs and external instance restarts. This is not an uninterrupted single-source run. The authoritative selection and checkpoint joins are in baseline/recording_audit.json. Failed/slipping suffixes are excluded from that selection. Historical failure reports and manifests remain distinguishable; some superseded raw binaries were deleted with the owner's authorization after delivery.
All 232,518 water particles remain represented in the accepted trajectory, including distant spilled parcels. Full-water stability, resolution convergence, wetting/capillary calibration, complete energy accounting and complete signed port-flux history are not certified. The historical missing entry flag for particle 119770 is not reconstructed. Metadata host paths are normalized to @TASK_ROOT@; numerical arrays and their hashes are unchanged.
Files and reproduction
baseline/archives/ contains per-case inputs, frozen source, accepted recorded frames, and preserved terminal artifacts. baseline/manifests/ provides per-file SHA-256 and sizes. These are published incrementally; use baseline/release.json to check whether publication is complete, then pin its repository commit SHA.
python -m pip install huggingface_hub hf_xet
python reproduce/fetch_reproduction.py --revision FULL_COMMIT_SHA --case all --destination ./artifacts
# After fetching, use the project MuJoCo/Warp environment for recorded-state replay.
# Set DROPPERLAB_ARTIFACT_ROOT to the artifacts directory.
python reproduce/render_actual_lineage.py --audit audit --out replay
The downloader verifies archive OIDs and every extracted file before exposing the local case. Rendering advances no physics and fabricates no fluid states. Glass and near-view arms are translucent only for visualization.
v1/ contains the developing successor: empty dropper on rack A, inverted empty tube on rack B, parallel pickup, stationary receiver after initial placement, and dropper return. It must not be confused with the delivered baseline or treated as completed.
Attribution
RoboDojo-derived assets retain the accompanying MIT attribution in licenses/RoboDojo-LICENSE.txt. MuJoCo, NVIDIA Warp, and other external runtime dependencies retain their own licenses and are not bundled as installed environments. Credentials, browser sessions and raw terminal logs are excluded.
Verified baseline archive storage (2026-10-13)
All 14 baseline archives were downloaded back onto the compute host and verified file by file: 7,381 members, SHA-256, byte counts, and checkpoint/recording NPZ readability checks. The verified immutable artifact revision is abc36e5ce48c07dcc527c3718e81d825437e8c86. Only then were the old on-host baseline recording frames and upload tarballs retired (39,315,252,148 bytes). Frozen source, scene inputs, checkpoints, reports, and the delivered video remain on the host; the developing V1 run is untouched. See baseline/storage_status.json. This is storage/reproducibility verification, not physical qualification.