Corrected shipwreck evaluations and historical demonstrations¶

CabinRecovery and CargoRelease have completed ACT/DP evaluations on 30 matched, independently reserved states per task (8000–8029). Together with the six-task core and HotStab, this gives nine tasks with reported ACT/DP evaluations. Different scene, data and observation protocols are kept separate; the original six-task macro-average is unchanged.

Cabin uses 80 fresh accepted demonstrations after the door, hinge and handle corrections, split into 76 training and four validation episodes. The collection retains all 160 attempts and excludes failed/audit-rejected trajectories from training. Historical collision-defective recordings are not reused. Expert command replay passes 12/12 for each task on development states.

Cabin's door/HDD–hinge geometry audits pass 27/30 ACT and 30/30 DP test episodes. Those audits inspect recorded 30 Hz frames with the unchanged 0.5 mm tolerance; they do not cover all contacts or physics substeps. Audit failures remain in the 30-episode denominator. Low policy success is a completed benchmark outcome, not evidence that the expert or scene has not been implemented.

Per-state outcomes and reports and the import manifest pin the evidence to handoff/openchest-base commit 3f8a5fb073362112977b237052612ea7511cc555. The evaluated runtime is df188e9b5e51e0cc6c7291c5eeb0e76499ce84b4. This PC checked archival consistency and presentation; it did not retrain or repeat these physical evaluations. Original reports retain their historical paths and links.

Reproduce and obtain artifacts¶

The publication release mirrors all four selected checkpoint files, the upstream evidence/video archives and their original SHA-256 values. Its runtime-source archive exports the exact upstream commit above without merging it into the frozen core runtime.

Extract wreck-corrected-v3-runtime-source.tar.gz under work/.maintenance/checkouts/ (or a fresh contained directory on another PC). Install the dependencies and licensed Reach Alpha meshes according to the snapshot's README and docs/openchest-second-machine.md, then follow the archived evaluation commands. Run from that extracted runtime's root, using the checkpoint filenames specified there. This source export is a reproduction input, not a new clean-machine replication claim. Raw RGB is not included or needed for checkpoint evaluation; training requires the original recordings or separately regenerated data.

Website and task status¶

Results → High-level policies → Shipwreck tasks contains the corrected curves and all four complete scored first-test-state recordings. Historical v1 tables and development diagnostics remain explicitly labelled in the appendix. Water-appearance films are expert recordings, not learned-policy robustness tests.

The tray-based Toss Ball at 3f8a5fb is historical. The wall-only v2 variant from f16b6cf is now included as Wall Toss: one paired expert example in still water and cross-current, with identical high-level commands. It appears in the paper and Physical effects section, without ACT/DP scores or inclusion in the nine evaluated model tasks.