FailBank Turns Runtime Shield Feedback into Persistent VLA Policy Gains (+25.4 Points Success)
notredame · hf · 2026-10-05
FailBank is a four-stage self-evolution framework that converts runtime feedback into persistent policy improvement for vision-language-action models, addressing the persistent policy-shield mismatch that pure runtime shields create.
- Collection: a fixed CBF-based safety module acts as an observe-only teacher, producing counterfactual corrections while the policy stays in control;
- Outcome-aware admission: useful proposals become corrective targets, successful uncorrected actions are kept as quiet anchors for guarded LoRA updates.
On the VLA-Arena benchmark across two VLA backbones: +8.5 and +6.9 points task success over base policies while cutting policy-induced cumulative cost by 35.6% and 23.8%; versus runtime shielding, +25.4 and +9.5 points at comparable cost. Runtime feedback can serve as persistent policy supervision rather than a temporary action constraint.
More from Embodied
- Reverse engineering Neuralink's new decoder architecture from a sparse blog post — melnykowycz · 2026-10-06
- Home robot reviews will be wild: 'Amazing gait, put my keys in the fridge' — VraserX · 2026-10-06
- EMPIRIC: an agent that extends physics engines with code for missing mechanisms — tomssilver · 2026-10-05
- Snap Staffers on Why $2,195 Specs AR Glasses Landed Chunky: The Meta Boogeyman — richardjnieva · 2026-10-05
- Panasonic building its own humanoid robot, targeting 2029 production in its own factories — CyberRobooo · 2026-10-05
- Reka releases Rho-1 research preview: a 19B omni model unifying text, images, video and robot actions — RekaAILabs · 2026-10-05