Refined Policy Distillation: From VLA Generalists to RL Experts

Tobias Jülg, Wolfram Burgard, Florian Walter

Abstract: Recent generalist Vision-Language-Action Models (VLAs) can perform a variety of tasks on real robots with remarkable generalization capabilities. However, reported success rates are often not on par with those of expert policies. Moreover, VLAs usually do not work out of the box and often must be fine-tuned as they are sensitive to setup changes. In this work, we present Refined Policy Distillation (RPD), an RL-based policy refinement method that enables the distillation of large generalist models into small, high-performing expert policies. The student policy is guided during the RL exploration by actions of a teacher VLA for increased sample efficiency and faster convergence. Different from previous work that focuses on applying VLAs to real-world experiments, we create fine-tuned versions of Octo and OpenVLA for ManiSkill2 to evaluate RPD in simulation. As our results for different manipulation tasks demonstrate, RPD enables the RL agent to learn expert policies that surpass the teacher's performance in both dense and sparse reward settings. Our approach is even robust to changes in the camera perspective and can generalize to task variations that the underlying VLA cannot solve.