DiMaS: Distribution matching steers robot behavior with fine-grained control

Robotic manipulation powered by vision-language-action (VLA) models has made leaps, but a critical gap remains: fine-grained control over *how* a robot performs a task. Traditional representation steering, a go‑to interpretability tool, shifts representations along fixed linear directions—a technique that works for language models but falters in visuomotor policies.

DiMaS tackles this by replacing linear movement with distribution matching. Instead of a single shift, it transports the entire representation distribution from one behavioral state to another. The approach is tailored to flow‑matching VLA architectures and has been validated on two leading models, showing consistent behavioral control even as training and evaluation tasks diverge.

The paper also explains why linear steering underperforms: in VLA models, behavioral features are linearly decodable (you can read them off) but not linearly steerable (you can’t change them by pushing in a straight line). This insight motivated DiMaS’s distribution‑matching design, which bypasses the nonlinear geometry of the action expert.

With open‑source code and demonstration videos, DiMaS offers a practical tool for researchers and engineers seeking to dial in robot behavior—like adjusting grasp force or motion style—without retraining the entire policy. This work paves the way for more interpretable and controllable robotic systems.