PDRVLA adds physical dynamics representations to robot control
PDRVLA is a Vision-Language-Action model whose lightweight Physical Dynamics Module extracts physical tokens from vision-language features to condition action generation without online video generation. The module predicts future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen pretrained video world model. Training uses publicly available data expanded through human-to-robot synthesis, subtask annotation, and embodied chain-of-thought annotation. The paper’s authors report success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv.
PDRVLA addresses the high latency that video world models can introduce into the control loop by using physical-dynamics representations without generating future videos online. The report also describes ablations showing improved future prediction with world-model supervision and examining the effects of subtask conditioning, future object-pose supervision, and embodied chain-of-thought training on action success.
Computer Science > Robotics
arXiv:2610.37381 (cs)
[Submitted on 27 Oct 2026]
Title:PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
Authors:Hanqing Wang, Zhaotian Cai, Ruichuan An, Bohan Zeng, Tao Zhang, Yilin Chen, Xu Jiang, Kaiquan Chen, Chongyang Wu, Haoran Li, Fan Yang, Mingan Lin, Wentao Zhang, Zenan Zhou, Youzheng Wu
View a PDF of the paper titled PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations, by Hanqing Wang and 14 other authors
#PDRVLA#Physical Dynamics Module#LIBERO#LIBERO-Plus#RoboTwin 2.0#SimplerEnv