Event
Proposes a visual failure attribution method for VLA models
1 reports · 1 sources
Carson Sobolewski et al. propose a decomposed risk model for failure attribution in vision-language-action (VLA) models: when multiple visual corruptions coexist, it determines which corruptions contribute to failure risk and estimates how much benefit their removal would bring. The method separately learns representations of the scene and of the corruptions, and estimates the removal benefit solely from the current image and task instruction via latent-variable intervention, without requiring paired clean images or additional counterfactual rollouts. The authors report that in simulations of OpenVLA and π0.5 and in real-world testing of π0.5-DROID on a Franka robot, the method outperforms monolithic baselines and vision-language model (VLM) baselines, with attribution error about one-third that of frontier VLMs; they also report that the method adapts to execution context and corruption intensity and can be used for online failure monitoring on hardware.
Report timeline
- Failure Attribution for VLA Models: Estimating the Benefit of Removing Visual Corruptions
The paper defines the problem of failure attribution for vision-language-action (VLA) models: when multiple visual corruptions coexist, determining which corruptions contribute to failure risk and how much benefit their removal would bring. The authors propose a factorized risk model that learns separate representations of the scene and of corruptions, and estimates removal benefits through latent intervention using only the current image and task instruction, without paired clean images or additional counterfactual rollouts.
arXiv cs.RO ·