Event
Study proposes safety requirements for federated VLA adaptation
1 reports · 1 sources
Vigneshwaran Subbaraju, Obaidullah Zaland, and Monowar Bhuyan propose that combining federated learning with vision-language-action (VLA) policies and sharing human corrections across tasks, environments, people, and robot embodiments carries safety risks: shared updates that improve average task success rates may suppress rare safety interventions or produce unsafe responses on another robot. They accordingly propose three requirements: preserve the context and uncertainty of multimodal corrections when deciding what to share; evaluate candidate updates in the closed loop of each receiving robot before deployment; and, when assessing the retention of safety-critical evidence, also assess its leakage. The authors state that the paper offers counterexamples and a falsifiable evaluation agenda rather than experimental results.
Report timeline
- How Human Corrections Travel Across Robots: Safety Issues in Federated VLA Adaptation
Regarding the prospects of combining federated learning with VLA policies, the research points out that sharing human-correction adaptations across tasks, environments, people, and robot embodiments carries safety risks: a shared update that raises average task success may suppress rare safety interventions or produce unsafe responses on another robot.
arXiv cs.RO ·