Skip to content

Latest commit

 

History

History
35 lines (30 loc) · 2.01 KB

File metadata and controls

35 lines (30 loc) · 2.01 KB

Boundary-Guided Replay Design Note

This file is a short design note for the current repository state. The active submission manuscript is paper/main.tex; result provenance is recorded in results/README.md; executable experiment entry points are listed in README.md.

Boundary-Guided Replay (BGR) trains decision policies from replayable states near an estimated success-failure boundary. For each replayable state, BGR estimates a recovery curve over perturbation radius, derives a critical radius, and samples training perturbations near that state-conditioned boundary while mixing in clean and broader coverage examples.

The current evidence is tiered:

  • Synthetic experiments support the intended recovery-margin mechanism, with a 15-seed rendered study and a 30-seed confirmation in the anonymous artifact.
  • Procedural grid-margin experiments support the main recovery-margin claim. The completed 30-seed full-baseline comparison and held-out 30-seed replication pool to 60/0 paired RAUC wins for BGR over uniform.
  • Robot-suffix experiments are positive manipulation-style evidence. The coverage-aware BGR-Suffix variant improves clean success, object RAUC, transfer RAUC, and AULC over uniform suffix replay across original and held-out 30-seed sweeps and four suffix stress regimes, while uniform remains higher on median critical radius.
  • OpenVLA/LIBERO experiments are audits of recovery-curve measurement, boundary selection, action-label/TFDS plumbing, and bounded full-goal evaluation. The compact OpenVLA validation artifact verifies 2,048-transition matched BGR/random exports with 7D actions and 8D state, but these audits are not positive BGR fine-tuning evidence because BGR does not stably outperform both matched random selection and the unadapted official checkpoint.

The paper therefore frames BGR as a boundary-centered replay principle with controlled procedural evidence, positive manipulation-style simulator evidence, and learned-policy audits, not as a completed robotics fine-tuning claim.