Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models

Can runtime safety feedback be converted into persistent learning updates that improve a VLA policy's success and safety trade-off, rather than only correcting individual actions?

VLA policies for manipulation must balance completing tasks with avoiding unintended contact in scenes with lookalike and protected objects. Runtime shields like AEGIS can correct unsafe actions at execution time, but they leave the policy unchanged, so repeated policy-shield disagreements can redirect the robot until it times out or enters unfamiliar states.

Before FailBank, runtime safety was typically handled by control barrier function shields with visual grounding, such as AEGIS, that project unsafe actions at execution time, by constrained learning methods such as SafeVLA, or by failure monitors. These approaches either leave the underlying policy fixed or require safety-specific training and cannot turn repeated intervention patterns into durable policy behavior.

FailBank uses a four-stage loop. A fixed CBF teacher observes rollouts and logs counterfactual corrections without changing executed actions, preserving the policy's own failure distribution. Outcome-aware admission keeps useful correction records and successful uncorrected actions as quiet anchors, weights them by trigger provenance and rollout outcomes, and accumulates them in a failure bank. Each round fits a fresh LoRA adapter from the same base checkpoint and accepts it only if held-out full-chunk flow loss and first-action drift stay within fixed limits; the accepted policy collects the next round and requires only RGB and proprioception at deployment.

Across ten Level 1 and Level 2 static-obstacle tasks, mean success rose from 66.0% to 74.5% for π0.5 and from 47.5% to 54.4% for π0, gains of 8.5 and 6.9 points. Policy-induced cumulative cost fell from 47.76 to 30.76 for π0.5 and from 10.62 to 8.09 for π0, reductions of 35.6% and 23.8%. Against runtime shielding with AEGIS, FailBank raised mean success from 49.1% to 74.5% on π0.5 and from 44.9% to 54.4% on π0, while policy-induced cost was 30.76 vs 29.01 and 8.09 vs 6.27, respectively. On held-out states of the collection task, π0.5 success rose from 71.1% to 93.3%; on unseen Level 1 tasks from 82.5% to 88.7%; and on harder Level 2 tasks from 50.5% to 59.4% without Level 2 update data. For π0, unseen Level 1 success rose from 51.0% to 62.7% and Level 2 from 35.7% to 40.0%. Ablations showed shield-in-loop collection produced shield-induced failure records for 88.9% of failure cases, and adding quiet anchors kept success at 89% while reducing policy-induced cost from 20.88 to 16.52 on one task.

The authors note the method currently targets continuous flow-matching policies and static obstacles, and would need temporal obstacle prediction for dynamic scenes. The evaluation is simulation-only on VLA-Arena, uses two flow-matching backbones, and does not report real-robot trials or compute cost. The guarded update also accepts or rejects on held-out flow metrics only, not on closed-loop task success.