FastOPD: On-Policy Distillation for Lightweight VLA Deployment

FastOPD. On-Policy Distillation for Lightweight VLA Deployment. By Yoojin Oh and nine others, from KAIST and Sungkyunkwan University. Published on the second of October. No code repository is linked to it.

Vision-language-action foundation models have grown quickly, and the growth has bought better manipulation and better generalization. It has also bought high computational cost, and that cost makes putting these models on real robots harder and harder. The usual remedies take one of two routes. Design a smaller architecture, or cut the number of iterative denoising steps that a flow-based policy needs to produce an action.

FastOPD takes a different route: distillation. It is a framework for going from a large foundation model to a lightweight one, and the transfer is done by efficient on-policy distillation. A compact student is trained to learn the dynamics of the large teacher, so that what gets deployed is the small model.

Two ingredients make up the training signal. The first is a flow map, adapted so that the teacher can supervise the student from a single state. The second is a self-consistency objective, combined with that supervision to shape the compact student. The authors also give a theoretical result. Minimizing this objective lets the distilled student recover a distribution on par with the one an ideal few-step teacher would induce. In other words, the student is not only smaller; in principle it stands in for a teacher that needed only a few steps.

The evaluation spans several foundation policies, in simulation and on real hardware. On LIBERO, FastOPD keeps eighty-four percent of the performance of pi zero point five while using only two inference steps. Inference latency drops by seventy-eight point one percent, and the average success rate beats existing few-step distillation baselines.

With LingBot-VLA as teacher, on RoboTwin 2.0, FastOPD raises the single-step success rate over the base student by fifteen point nine percentage points. The method also carries over to a World Action Model, and a compact student distilled from MolmoAct2 is deployed on a real robot.

The pitch is that a robot does not have to run the big model to behave like it. A student trained on the teacher's own dynamics, with a guarantee tying it to an ideal few-step teacher, keeps most of the performance at a fraction of the latency, and the same recipe reaches past vision-language-action models to a world action model too.