Steering VLA robot models at inference time lifts grasp rate from 14% to 74% without retraining

drmapavone · x · 2026-09-23

A new paper asks whether robot foundation models (VLAs) can be steered without retraining.

Drawing on control-theoretic notions of observability and controllability, the authors show state- and action-relevant information can often be linearly predicted from VLA internal representations
These representations can be steered at inference time to change robot behavior, with no fine-tuning
On a real robot, the approach raised the preferred handle-grasp rate from 14% to 74% with only 1% inference overhead
The broader idea: representation-level control as a lightweight interface for adapting robot foundation models to new preferences and constraints

Joint work with Hugo Buurmeijer, Carmen Amo Alonso and Swann Aiden; paper is publicly available.

More from Embodied