PDRVLA adds physical dynamics tokens to VLA models, topping LIBERO with 99.3%
Hanqing Wang · hf · 2026-10-28
PDRVLA is a vision-language-action model that augments a VLM backbone with a lightweight physical dynamics module, supervised by a frozen video world model to predict future visual features—avoiding online video-generation latency at inference.
The architecture feeds multi-view images and task instructions through a VLM; intermediate features pass through the dynamics module into physical tokens, which are combined with final features in an action expert to emit 8-action chunks. Training-only supervision uses direct RGB/depth/optical-flow prediction plus auxiliary world-model supervision.
- LIBERO: 99.3%, beating π₀.5, Cosmos-Policy, Xiaomi-Robotics-0, Qwen-RobotManip, and InternVLA-A1.5
- LIBERO-Plus: 89.9%, ahead of Qwen-RobotManip (89.0%)
- RoboTwin 2.0: 93.6%, edging out LingBot-VLA 2.0 (93.2%)
- SimplerEnv: 81.5%, above InternVLA-A1.5 (80.8%)
More from Embodied
- Dyna Robotics unveils Taku robot with a craftsman-first video, no spec sheet — heyshrutimishra · 2026-10-28
- Alibaba DAMO releases RynnWorld-Latent: cross-embodiment world model pretrained on 553.5M frames — Alibaba-DAMO-Academy · 2026-10-28
- Developer hooks GPT into a Hasselblad: focus, shutter and chat, open-sourced — rrRadium · 2026-10-28
- Robots Coming for Your Jobs: A Playful Reaction to a Demo — alysha_lobo · 2026-10-28
- New benchmark tests whether home robots navigate more cautiously around babies than trash bins — SeoulNatlUniv · 2026-10-28
- Early adopter revisits AI hardware: still meh, runs hot without opening apps — alexmacgregor__ · 2026-10-28