Can we steer a robot foundation model without retraining it? Our new paper explores this question through control-inspired notions of feature observability and controllability in Vision-Language-Action (VLA) models. We find that state- and action-relevant information can often be linearly predicted from VLA internal representations—and that these representations can be steered at inference time to change robot behavior, without fine-tuning. As an example, on a real robot, our approach increased the preferred handle-grasp rate from 14% to 74%, while adding only ~1% inference-time overhead. The broader idea is that representation-level control could provide a lightweight interface for adapting robot foundation models to new preferences and constraints—without retraining. Joint work with Hugo Buurmeijer Carmen Amo Alonso @SwannAiden Paper: #PhysicalAI# #Robotics# #VLA# #RobotLearning# @StanfordEng @StanfordAILab