ModAR challenges RGB representations, beating 6B video models with 30M parameters

Researchers propose ModAR, the first world-action model that autoregressively denoises multiple future modalities—point trajectories, DINO features, depth maps—before predicting actions. Trained from scratch with just 30M parameters, it outperforms fine-tuned 6B video models while using 20x less compute, challenging RGB-based representations.

2026-09-16 ~ 2026-09-17 · 3 related posts

• ModAR: First Multimodal Autoregressive World-Action Model Cuts Training Compute 20x — Adam Hung · 2026-09-16
• ModAR: a 30.1M robot world model trained from scratch beats a 6B video-model baseline — CSProfKGD · 2026-09-17
• ModAR predicts robot futures one modality at a time; 30M model beats 6B video model — k7agar · 2026-09-17