Faster World Action Models for Robotics

open-gigaai · hf · 2026-07-16

This work presents improvements to the **World Action Model (WAM)** in robot policy learning. The original WAM explicitly generates future videos during inference, which aids action learning but incurs massive overhead, hindering real-time closed-loop control. The authors propose **GigaWorld-Policy-0.5**, decoupling training and inference: - **Training Phase**: Combines Action-Conditioned World Modeling with WAM training to tightly couple visual dynamics with action representations - **Inference Phase**: Performs action decoding only, eliminating future video generation to reduce overhead - Introduces **Mixture-of-Transformers** to separate visual dynamics modeling and action generation into specialized experts, boosting inference efficiency - Employs an **agent-based AutoResearch** pipeline to automatically search training configurations, minimizing manual tuning The authors report that this method achieves an action inference latency of **85ms** on a local **RTX 4090**. Experiments and ablations demonstrate enhanced real-time robot control while retaining the training benefits of future visual dynamics.

Original post →

More from Embodied

Embodied channel →