Peking University and JD's PDRVLA Hits 99.3% on LIBERO With 57M-Param Physical Dynamics Module

Peking University and JD.com's PDRVLA bolts a 57.3M-parameter Physical Dynamics Module (1.4% of a 4B backbone) onto a Qwen3.5-4B VLA, learning 50 physical tokens via future RGB, depth and optical-flow prediction. The model scores 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv, beating Qwen-RobotManip, InternVLA-A1.5 and π₀.₅ on all three while using only publicly available training data.