Skip to content

Add PiPER contact history v0.3 and publish the 84-attempt comparison - #13

Open
Tarzelf wants to merge 5 commits into
feat/piper-grip-recoveryfrom
feat/piper-contact-memory-v03
Open

Tarzelf wants to merge 5 commits into
feat/piper-grip-recoveryfrom
feat/piper-contact-memory-v03

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 12, 2026 •

Copy link
Copy Markdown
Contributor

PiPER's previous fixed drift guard aborted grasps that the original controller completed. This adds a causal contact-history controller with a bounded closed hold and reassessment, alongside the same acquisition probe without postqualification intervention so the comparison separates probing from intervention.

The frozen confirmation retains all 84 native attempts across three frictions, two masses, two shared layouts and seven variants per scene. Every placement has its own complete requested-arm/open-jaw control. All failures, command/state trajectories, timing summaries, maxima and miss counts, source identities and original model notices remain public; development evidence is separate. Native mechanics, gains, force caps and the task scorer are unchanged.

Fresh confirmation: history 6/12, original 4/12, probe-only 4/12; all 84 attempts valid and all 48 negative controls 0. Both additional successes are 60 g blocks at friction 0.8. Five closed holds/resumes across four completed history attempts include one task requiring a second hold; six acquisition retries still fail. Target scaling does not imply immediate speed reduction. In moderate-friction 40 g cases the new intervention adds about 2.9 simulated seconds without additional success.

Validation: 35 controller/native tests, 22 existing benchmark regressions, 11 new benchmark/receipt tests, 20 publisher tests and 19 preparer tests passed. Installed-wheel sources matched all 11 frozen runtime hashes before an OS network-denied confirmation run. Publication preserves exact sealed bytes and independently checks decoded evidence; file integrity does not establish physical competence.

This is an authored simulator controller using privileged state. It does not train a policy from footage or qualify a physical robot. Version v0.3 names this experiment; the package version is separate.

Complete confirmation evidence · Local run guide · Robot Simulation Lab · Research article

Verified Hugging Face revision: a6bf57fd64e61486dafaa93fd96168f953ccff80; all 203 public files match the reviewed plan.

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