PDRVLA conditions robot actions on physical dynamics without video generation
Original title: PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
It reports 99.3% success on LIBERO and 81.5% on SimplerEnv using public data and a frozen world model.
PDRVLA is a vision-language-action model whose Physical Dynamics Module extracts physical tokens from vision-language features to condition robot actions without generating future videos online. The module predicts future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen pretrained video world model. Training uses only public data plus human-to-robot synthesis, subtask annotation, and embodied chain-of-thought. On four simulation benchmarks it reports 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv.
- Reported results are simulation success rates, not physical-robot trials.
- Ablations credit world-model supervision with better future prediction.
- Subtask conditioning, future object-pose supervision, and ECoT training are evaluated.
- The module is meant to avoid online video-generation latency in the control loop.
From the original article
1 min read at huggingface.co
Vision--Language--Action (VLA) models generate robot actions from visual observations and language instructions, but their understanding of physical dynamics remains limited. Video world models can capture scene dynamics, yet generating future videos introduces high latency into the control loop.
By Hanqing Wang, Zhaotian Cai, Ruichuan An · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.