Skip to content

Improve PiPER placement reliability with bounded re-probe and alternate grasp - #15

Open
Tarzelf wants to merge 3 commits into
feat/piper-contact-recovery-v04from
feat/piper-reliability-v05
Open

Tarzelf wants to merge 3 commits into
feat/piper-contact-recovery-v04from
feat/piper-reliability-v05

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 12, 2026 •

Copy link
Copy Markdown
Contributor

The earlier controller could leave a stable suspended grip below its resume-height gate, while some heavier blocks could not be lifted with its original side pinch. This adds one bounded second lift and a fixed diagonal grasp planner, preserving native mechanics, force/gain limits, task scoring and the full 20-second budget.

Three protocols and runtime e76b2e4 were committed before fresh execution. Across 36 matched placements, v0.4 completes 15, second-lift recovery completes 18, and the full alternate-grasp candidate completes 23: eight gains, no lost baseline successes. All 252 attempts are valid, all 144 negative controls fail, and no prohibited contacts or solver warnings occur. Thirteen full-candidate failures remain, including a dropped-near-target block correctly rejected by retained-carry scoring.

The full candidate is 0.772 simulated seconds slower on its fifteen shared successes. All three second probes qualify and complete in the lift-only ablation; the full candidate triggers none, so its measured gains come through the fixed grasp/planner route. Loaded-host timing retains 1,035/252,000 software windows over 20 ms and a 3.082-second maximum. This is authored privileged-state simulation, not a learned skill or physical real-time qualification.

Validation: 131 relevant native/controller/benchmark tests and 52 publication tests pass. All 24 historical baseline reproductions preserve exact commands, qpos/qvel, metrics and outcomes. Every fresh receipt passes portable consistency checks and full native FK/policy/IK reconstruction. The installed wheel runs with OS network denial and all sixteen source/lock hashes verified. All 200 development attempts, including rejected collision-producing candidates, are retained separately from confirmation. Long CI simulation comparisons are split into separate jobs to avoid the earlier combined-job timeout.

Complete results and reproduction · Research article · Robot Simulation Lab

All 563 public Hugging Face files (429,121,291 bytes) were downloaded and matched at immutable revision d10b2c8. The outer evidence archive includes that receipt and the 28 passing GitHub checks for code/results commit 0545ac4; later publication/documentation commits have their own checks. Public and local Space checks pass at desktop, 390px and 320px widths without browser errors or page overflow.

The roadmap also defines a planned reusable asset library for arms, grippers, objects, environments and sensors, starting with shoe organization. It separates visual coverage from measured contact/deformation evidence and binds assets to compatible skill benchmarks. This catalog is planned work.

The research article is live. All 312 added/changed public site files match the tested local bytes, including all 270 sealed native evidence files. Live research and Hugging Face pages pass desktop, 390px and 320px browser checks with no page overflow or JavaScript errors. All 28 checks on final training head 05d267f also pass.

Stacked on #14; the training changes remain a public proposal on this branch. Research PR #30 is merged and deployed.

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