Skip to content

Wall Toss: one matched expert-command pair

The wall-only v2 task is an recording of current sensitivity, not a tenth ACT/DP benchmark. A 50 mm, 100 g capsule passes through the wall in still water (success at 7.80 s); a +Y 0.2 m/s current causes a rim strike under identical high-level commands. Both 324 × 8 command arrays match byte for byte.

The website's Physical effects section presents full paired movies, three camera choices and recorded lateral-position curves. Task cards retain clean expert movies; learned-policy outcomes remain in Results.

Evidence and runtime

Fetch artifact wall-toss-runtime using scripts/fetch_paper_artifacts.py and extract it into a fresh maintenance directory. Follow that runtime's installation requirements, including authorized Reach meshes, then use the commands in the full task contract. Never extract it over a pinned core, HotStab or bimanual runtime.

The comparison changes the current acting on both robot and capsule. It does not isolate capsule drag alone. Matching low-level feedback activity does not imply identical realized forces. The source also contains separately labeled contact-free calculations; they are not extra simulated task successes.

Selected expert cameras

Fresh CabinRecovery and HotStab external/base/gripper recordings are imported from the same upstream commit. These are presentation episodes with their own sampled audits, not replacements for benchmark trajectories. The website retains our completed PullLever/PushSlider experts and core films: the upstream branch's zero-action previews and older core presentation versions are not substituted.

The sphere model and optical variants are uncalibrated simulation approximations. This update introduces no new training, random-reset success rate or hardware claim.