Megadose The AI curve, compressed.

JointGV: Structure-Aware Correction of Visual World Models with Parallel Graph Dynamics for Long-Horizon Modeling

· ArXiv · AI/CL/LG ·
An actor trained on JointGV-corrected predictions completed 46 of 50 simulated pendulum-control episodes; one trained on the original predictions completed none.

JointGV runs a scene graph alongside a visual world model, using predicted changes in object positions and relationships to correct drifting visual predictions. The plug-in leaves the original visual dynamics untouched by graph-learning and correction gradients. Across 32-step benchmark rollouts, the authors report joint-position error reductions of 11–28%. On real-robot pendulum sequences, it cut DINO-WM’s 32-step pole-angle error from 12.56° to 6.69°.

ArXiv · AI/CL/LG's note

score 5

Categories: Research