---
license: other
library_name: pytorch
tags:
- robotics
- action-chunking
- se3
- b-spline
- minimum-jerk
- robotwin
datasets:
- Shiki42/up-vla-precision-recovery-15k
---

# UP-VLA Structured Recovery Checkpoints

Protocol-registered recovery experiments:

- E001: endpoint/direct/residual H16 action chunks; completed inconclusive.
- E002: same-split endpoint and residual H16 baselines plus clamped cubic and true quintic minimum-jerk SE(3) residual heads; completed rejected. The selected minimum-jerk head reduced fresh-test grouped jerk by 18.50% but missed the frozen 25% effect gate and worsened path score by 4.18%, above the 2% bound. Real Pi0.5 rollout is not authorized.

The source dataset includes RLBench-derived material with non-commercial research restrictions. Treat these checkpoints as research-only unless all upstream license obligations have been independently resolved. See each experiment card and immutable run/evaluation receipts.