Skip to content

Add graspability controls and expose shoe/gripper model limits - #20

Open
Tarzelf wants to merge 2 commits into
feat/continuous-shoe-grip-v02from
feat/graspability-v03
Open

Tarzelf wants to merge 2 commits into
feat/continuous-shoe-grip-v02from
feat/graspability-v03

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 15, 2026

Copy link
Copy Markdown
Contributor

The prior shoe task combined retrieval, wrist-tool interference and grip retention, so its failures did not identify whether the object or simulated gripper was limiting pickup. This adds a separate CPU direct-grasp diagnostic with three rigid shapes, explicit 30 g / 300 g loads, mounted/absent rake conditions, and an interactive 3D evidence viewer.

The controller and source were frozen at 73057dc259c6a19344906f4a43a283553e17ade0 before one fresh 24-attempt evaluation. All runs are numerically valid. Lightweight outcomes vary by shape and layout; every 300 g attempt fails. Removing the rake gains one narrow-proxy lift and loses one block lift. These are diagnostic conditions, not a shoe success rate or a comparison against the earlier retrieval task.

The independent scorer verifies initial settlement and every native 2 ms tick of the final 500 ms for relative rise, bilateral contact and absence of other support. Complete failed trials remain available. Publication rechecks the full traces and preserves earlier releases. CI now builds the new viewer, runs development controls and validates exported evidence.

Validation: 48 native/regression tests, 26 publisher tests, independent verification of all 24 compiled models and scores, and browser loading/final-frame inspection of all 24 replays. The recorded host ran 158.4 simulated seconds in 8.404 native wall seconds, with 20.032 seconds of separate audit/export work.

Limitations are explicit: authored poses and privileged state, rigid solid shoe proxies, uncalibrated drive force, and soft joint-limit overtravel. Both successful lightweight full-shoe cases exceed nominal 70 mm jaw span during retention. Commands remain bounded, but hardware travel/force equivalence is not established. The frozen scoring criteria and results have not been changed after evaluation.

  • Research
  • Recorded comparisons
  • Reproduction: docs/shoe-graspability.md
  • Complete counts, hashes and portable audits: benchmarks/shoe-graspability-20261014

Stacked on feat/continuous-shoe-grip-v02 (PR #19); earlier experiment files and results remain intact.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant