Vision-language-action (VLA) models are designed to transfer knowledge from large-scale pretraining into robot control. However, a new arXiv preprint, VLA-Retain, points out that their manipulation performance remains sensitive to changes in observation conditions. The abstract frames this fragility as an open problem that recent work has started to address.
The paper introduces a benchmark for studying the cross-dimension effects of vision tuning in VLAs. The title suggests a focus on whether models retain useful capabilities after vision-side adjustments, and the benchmark is positioned as a way to measure those effects systematically. The available abstract excerpt does not include experimental results or detailed methodology.
Because this article draws on a single source, there are no independent findings to compare. The contribution, as described, is the benchmark itself and the framing of observation-condition sensitivity as a key issue for VLA manipulation.