Paper review · October 29, 2026
PDRVLA: A VLA model with physical dynamics Representation
PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
Physics tokens, trained to predict future RGB, depth and optical flow from intermediate VLM features, are fed into action generation. Without online video generation, this achieved high success rates on four simulation benchmarks.
Problem
VLA models have a weak physical understanding of how objects move and change during interaction. Generating future frames with a video world model can capture scene dynamics, but generating video inside the control loop adds large latency.
Method
A lightweight Physical Dynamics Module extracts a fixed number of physics tokens from intermediate VLM features. The action expert uses these tokens together with the VLM's final-layer features to generate actions. The physics tokens are trained to directly predict latent features of future RGB, depth and optical flow, and a frozen pretrained video world model conditioned on these tokens provides auxiliary supervision. The video generator is not run at deployment. All training data is public. It combines real-robot, UMI and simulation Trajectories with VL data, and adds synthetic data that converts EgoDex human videos into demonstrations on 12 types of robot arms (about 0.9M Trajectories), subtask labels and ECoT annotations.
Results
- Reported success rates are 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0 and 81.5% on SimplerEnv.
- In ablations, auxiliary supervision from the world model improved future prediction.
- The authors also separately analyzed how subtask conditioning, future object pose supervision and ECoT training affect action success rates. Numerical comparisons with baseline models are not included in the provided text.
Limitations
All evaluation is on simulation benchmarks, so real-robot performance is unverified. Comparison numbers against existing VLAs and detailed ablation numbers are cut off in the provided text. The data is large in scale, making it hard to fully separate the contribution of physics tokens from data effects. Automatic subtask and ECoT annotations may contain VLM labeling errors.
Why it matters
It is a practical compromise that brings a world model's ability to 'look ahead' into a robot policy without inference latency. It also documents in detail the pipeline for building large-scale cross-embodiment training data from public data only and for converting human videos into robot data, which makes it a useful reference for reproducibility.
Key terms
- VLA(Vision-Language-Action)
- A model that takes camera images and language instructions and directly outputs robot actions.
- Physics tokens
- Compressed tokens extracted from VLM features and trained to hold information about future appearance, geometry and motion. They serve as an extra condition for action generation.
- ECoT(Embodied Chain-of-Thought)
- Reasoning annotations for robots that spell out the scene description, task progress and next action in language.
- UMI
- A way of collecting demonstrations with a handheld gripper and a wrist camera. It yields end-effector Trajectories that are not tied to a specific robot.
Original
Authors: Hanqing Wang, Zhaotian Cai, Ruichuan An, Bohan Zeng, Tao Zhang, Yilin Chen, Xu Jiang, Kaiquan Chen, et al. · Posted 2026-10-27 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.