A method that turns runtime failure signals into new training records reported higher average task success and lower policy-induced cumulative cost than the corresponding base policies in a ten-task benchmark. For π0.5, mean success rate was 74.5%, compared with 66.0% for the base policy. For π0, it was 54.4%, compared with 47.5%. Mean policy-induced cost was 30.76 versus 47.76 cost units for π0.5, and 8.09 versus 10.62 for π0. Those differences corresponded to reported cost reductions of 35.6% and 23.8%.
Called FAIL BANK, the framework is built around a four-stage process. It logs counterfactual control-barrier-function, or CBF, proposals during policy-controlled rollouts. These are suggested corrections to the action being evaluated, rather than the action recorded by the policy. The system admits corrective records and quiet anchors, accumulates them in a bank, and accepts fresh LoRA adapters, targeted updates to the existing policy, only after a held-out guard. The underlying question is whether runtime evidence can become lasting supervision for later policy updates.
The benchmark and its safeguards
The evaluation used VLA-Arena's static-obstacle safety suite. The suite contains five tasks at each of three difficulty levels; the analysis covered Levels 1 and 2, collected training records on the Level 1 mango task, and did not put Level 2 update data into the main update. Reported benchmark outcomes included task success rate, official cumulative cost, policy-induced cumulative cost, and a score comparing each result with its base policy.
Candidate updates were not accepted simply because a benchmark score improved. The held-out full-chunk flow-loss ratio had to be no more than 1.10, and first-action drift, measured as mean absolute difference per coordinate, had to be no more than 0.05. Benchmark success rate and official cumulative cost were not used for acceptance.
The comparison is a trade-off
Against AEGIS, FAIL BANK reported a higher mean success rate on both model variants: 74.5% versus 49.1% on π0.5, and 54.4% versus 44.9% on π0. Its mean policy-induced cost, however, was also higher: 30.76 versus 29.01 on π0.5 and 8.09 versus 6.27 on π0. The result is a trade-off in the reported averages, not a universal win across both measures.
Some gains appeared outside the collection task
Reported success gains also appeared in held-out and unseen settings. On held-out Level 1 mango states, π0.5 moved from 71.1% for the base to 93.3% with FAIL BANK. Across four unseen Level 1 tasks, it moved from 82.5% to 88.7%, while the Level 2 average moved from 50.5% to 59.4%. For π0, the corresponding change was 51.0% to 62.7% across four unseen Level 1 tasks and 35.7% to 40.0% on Level 2.
The broader comparisons were less conclusive than the headline averages suggest. The primary significance test was a preregistered paired sign test. For π0.5 on the four unseen Level 1 tasks, the 82.5% to 88.7% change contained 51 wins, 45 losses and 104 ties, with p=0.61. The π0 Level 2 gain had p=0.13. These results leave the consistency of the out-of-sample gains unresolved.
The learning signal mattered
An ablation of the data-collection interface produced a different set of records when the shield was placed in the loop. Across 46 paired Level 1 mango cells, shield-in-loop collection turned eight otherwise successful rollouts into failures and rescued five failures. It left 37 successful episodes, compared with 40 under observe-only collection, and 88.9% of shield-in-loop failure records were shield-induced. The comparison was paired, not a randomized trial, so it describes a difference in collection behavior rather than a causal estimate.
On Level 1 onion, correction-only training was associated with a rise in success rate from 65% to 89% and a drop in policy-induced cost from 38.61 to 20.88. Adding quiet anchors left success at 89% and reduced cost further to 16.52. A separate five-round test showed why more accumulated data is not automatically better: the bank grew from 3.7k to 22.7k records, success peaked at Round 2, and policy-induced cost reached its minimum at Round 3. Later values fluctuated, although all five rounds remained better than base on both axes.
The hardest test remained open
Transfer to dynamic obstacles was not established. On 18 matched Level 1-T3 states, success was 55.6% for the base and 72.2% for the static-trained adapter. On Level 1-T4, both were 5.6%, and the pooled paired test gave p=0.5078. That result limits how far the static-obstacle findings can be extended.
The document is an arXiv preprint, version 1, dated 30 September 2026, and says code is available at Mingyuee88/FailBank. Its reported evidence is from the static-obstacle benchmark settings described above; transfer to dynamic obstacles remains unestablished.
Paper data and sources
Original title: Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models
Authors: Mingyue Cui, Zheyuan Liu, Yihan Zhu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-30
DOI: Not available
Original paper · Full text