Event

Proposes BAS-VLA, a task-semantic action calibration framework for VLA

1 reports · 1 sources

Shuaijun Liu et al. propose BAS-VLA, a task-semantic action calibration framework built on top of a frozen foundation VLA, to address two types of failure: unnecessary action drift produced by the policy when task semantics are unchanged but appearance changes, and the policy still following the original trajectory when key semantics such as the target object or constraints are changed. On four target-object replacement tasks, BAS-VLA raises the success rate on semantic-change tasks from 2.3% to 68.5% while maintaining performance on clean and semantics-preserving scenarios; on validated style-preserving shifts, the success rate rises from 42% to 70% without degrading clean performance. BAS-VLA by default adopts a calibration core centered on semantic disruption, supplemented by a semantics-preserving auxiliary pathway with selective evidence gating, correcting appearance-induced drift when task-preserving evidence is sufficient.