University of Geneva's GAR Speeds Up Robot AI Policies by 1.91x

Flow matching powers robot-control policies from Xiaomi Robotics and NVIDIA GR00T. These models can represent several plausible actions for the same state, a distribution that a simple Gaussian often approximates poorly. Researchers at the University of Geneva and University of Tübingen report a narrower result: after conditioning on an observation, action chunks sampled from these two policies closely match full-covariance Gaussians. Their study, Gaussian Distillation of Flow Policies ( paper ), turns that finding into a faster inference method.

Gaussian Action Readout, or GAR, freezes a trained flow policy and adds lightweight readouts for the Gaussian mean and covariance. The readout reduces iterative flow integration to one backbone pass. On 50 simulated manipulation tasks, GAR trails its teachers by 0.60 to 1.16 percentage points while accelerating model inference by 1.57× to 1.91×.

Why flow sampling taxes the control loop

A flow policy predicts an action chunk, a short sequence of future actions, by transforming random noise through a learned velocity field. Sampling requires several transformer evaluations as a numerical solver follows that field toward a clean action. Repeating those evaluations at every control step raises latency and limits control frequency.

Likelihood-based tools incur another cost. Offline reinforcement learning, safety filters, and imitation-learning metrics may need to score an action under the policy. A flow model generally computes that likelihood by integrating along a trajectory. A Gaussian provides the log-probability directly from its mean and covariance.