Vision-Language-Action Models: Adaptive and Uncertainty-Gated Policy Distillation for Robotic Manipulation

• Nicolas Mello USP

DOI: https://doi.org/10.5753/eniac.2026.28292

Resumo

The application of transformer-based models to robotics has led to Vision-Language-Action (VLA) models: networks that take camera images and natural-language instructions as input and output robot actions directly, and that generalize remarkably well across tasks. Their large size relative to conventional approaches, however, makes them costly for real-time control and hard to parallelize. A novel improvement is to distill the knowledge of a large VLA into a small expert policy invoked per task, a technique known as Refined Policy Distillation (RPD). RPD itself has two limitations: it weights the teacher’s influence equally throughout training regardless of how competent the student has become, and it trusts every teacher prediction equally even when the teacher is uncertain. We address both limitations with an adaptive schedule that reduces the teacher’s influence as the student improves, and an uncertainty-gated loss that discounts unreliable teacher predictions.

Referências

Black, K. et al. (2024). π0: A vision-language-action flow model for general robot control. arXiv:2410.24164.

Black, K. et al. (2025). π0.5: A vision-language-action model with open-world generalization. Physical Intelligence, arXiv:2504.16054.

Ghosh, D., Walke, H., Pertsch, K., et al. (2024). Octo: An open-source generalist robot policy. In Robotics: Science and Systems (RSS).

Julg, T., Burgard, W., and Walter, F. (2025). Refined policy distillation: Distilling task-specific expert policies from vision-language-action models. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

Kim, M. J. et al. (2024). OpenVLA: An open-source vision-language-action model. In Conference on Robot Learning (CoRL).

Liu, H. et al. (2025). VLA-RFT: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators. arXiv:2510.00406.

Mittal, M. et al. (2025). Isaac Lab: A GPU-accelerated simulation framework for multi-modal robot learning. NVIDIA, arXiv:2511.04831.

Nasiriany, S., Maddukuri, A., Zhang, L., Parikh, A., Lo, A., Joshi, A., Mandlekar, A., and Zhu, Y. (2024). RoboCasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS).

O’Neill, A. et al. (2024). Open X-Embodiment: Robotic learning datasets and RT-X models. In IEEE International Conference on Robotics and Automation (ICRA).

Pezzato, C., Ferrari, R., and Corbato, C. H. (2020). A novel adaptive controller for robot manipulators based on active inference. IEEE Robotics and Automation Letters, 5(2):2973–2980.

Ross, S., Gordon, G., and Bagnell, D. (2011). A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 627–635.

Rusu, A. A. et al. (2016). Policy distillation. In International Conference on Learning Representations (ICLR).

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017). Proximal policy optimization algorithms. arXiv:1707.06347.

Tao, S., Xiang, F., Shukla, A., et al. (2025). ManiSkill3: GPU parallelized robotics simulation and rendering for generalizable embodied AI. In Robotics: Science and Systems (RSS).

Wang, S.-C., Xiang, T.-Y., Zhou, X.-H., Gui, M.-J., Xie, X.-L., Liu, S.-Q., Wang, S.-Y., Jin, A.-Q., and Hou, Z.-G. (2025). VLA model post-training via action-chunked PPO and self behavior cloning. arXiv:2509.25718.

Wei, K. et al. (2025). πRL: Online RL fine-tuning for flow-based vision-language-action models. arXiv:2510.25889.

Xu, C., Li, Q., Luo, J., and Levine, S. (2024). RLDG: Robotic generalist policy distillation via reinforcement learning. arXiv:2412.09858.