Skip to content
InfoResearchPreprint

Robustness as a Transferable Skill: Cross-Task Adversarial Defense for Vision-Language-Action Policies

Published
Record updated
View JSON

Summary

This research studies whether robustness learned on source manipulation tasks can protect Vision-Language-Action (VLA) policies on held-out target tasks without target-side adversarial training. The authors introduce Source-Norm Transferable Robustness Vector (SN-TRV), which averages the direction of source defense updates and calibrates its magnitude with source-update norms, then applies it to the target visual encoder. Across three VLA backbones and fixed and adaptive attacks it improves robustness over naive averaging, and in physical PiPER trials average success under printed adversarial patches rises from 20.3% to 45.8%.

Mitigation

The source proposes applying the SN-TRV update to the target visual encoder, which requires no target adversarial data, target-side gradient optimization, or additional inference modules.