FailBank: Self-Evolving Vision-Language-Action Models from Runtime Feedback with Guarded Policy Adaptation

核心背景与行业痛点

Vision-Language-Action (VLA) models endow physical manipulators with end-to-end multi-modal policy control. However, practical deployment requires strictly avoiding unintended contacts, joint over-extensions, and physical collisions. The predominant defensive engineering pattern wraps policies in runtime safety shields—such as Control Barrier Function (CBF) filters—that override unsafe actions on the fly. However, external shields treat transient symptoms without adapting underlying neural weights. Unaware of safety limits, the VLA repeatedly generates violating trajectories, creating chronic policy-shield mismatches that paralyze long-horizon workflows through freezing or violent mechanical chatter.

架构亮点与底层机制

Researchers from the University of Notre Dame present FailBank (arXiv:2609.39820), a four-stage self-evolution framework:

1. Observe-Only Counterfactual Teacher: A deterministic CBF module monitors rollouts in an observe-only capacity, deducing counterfactual corrections in parallel while the baseline policy preserves operational agency.
2. Outcome-Aware Admission: Scrutinizes intercepted corrective proposals against holistic episode success, admitting only verified progress-enhancing interventions as supervisory targets while discarding spurious over-conservative overrides.
3. Guarded LoRA with Quiet Anchors: Preserves uninterrupted successful segments as quiet anchors within loss functions, confining LoRA parameter updates strictly to corrected boundaries to eliminate catastrophic forgetting of dexterity.

权威 Benchmark 与实测跑分对比

Evaluated on the rigorous VLA-Arena benchmark across multiple task difficulty tiers and two VLA backbones:

1. Superior Joint Success-Cost Operating Point: Improves task success rates by 8.5 and 6.9 percentage points while slashing cumulative policy safety costs by 35.6% and 23.8% across the two evaluated VLA backbones.
2. Catapults Success Over Static Runtime Shielding: Surges task completion rates by 25.4 and 9.5 percentage points relative to static CBF shielding pipelines while maintaining equivalent physical safety limits.
3. Converts Transient Constraints into Persistent Intelligence: Empirically proves that runtime interventions provide foundational supervision for continuous policy self-evolution rather than functioning merely as brittle execution throttles.

开发者实战落地与开箱指南

FailBank code and policy adapters are available on GitHub (Mingyuee88/FailBank). Robotics developers deploying manipulators across manufacturing, warehouse, or medical domains can deploy FailBank to ingest runtime barrier interventions into policy weights. By transforming runtime collision guards into self-evolving data loops, teams can graduate robots from fragile external overrides to intrinsic, safe spatial reasoning.