AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 15:18:24 UTC.

Briefings

A 2026 study introduces FailBank, a four-stage framework for Vision-Language-Action (VLA) models that converts runtime feedback into persistent policy improvements. Evaluated on VLA-Arena across two difficulty levels and two backbones, FailBank increased task success rates by 8.5 and 6.9 percentage points while reducing policy-induced cumulative costs by 23.8% and 35.6% respectively. Compared to runtime shielding, FailBank raised success rates by 9.5 and 25.4 percentage points without increasing costs, demonstrating runtime feedback as persistent policy supervision.

Key points

• FailBank is a four-stage self-evolving framework that converts runtime feedback into persistent policy improvements for VLA models.
• FailBank improved task success rates by 8.5 and 6.9 percentage points while reducing policy-induced cumulative costs by 23.8% and 35.6% respectively on VLA-Arena benchmark.

Context and limitations

Results are limited to two difficulty levels and two VLA backbones; specific failure modes and long-term generalization across diverse environments remain untested.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 08/10/2026, 14:01:25 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifgbgfx7wusrcqx6sh7etcxe5hmfghngfy3kibt5fq5d4omkd6jyu