Coupled Local and Global World Models for Efficient First Order RL
Researchers have developed a new method for training reinforcement learning (RL) policies within learned world models, bypassing the need for traditional simulators. This approach utilizes a decoupled first-order gradient (FoG) technique, combining a full-scale world model for accurate trajectory generation with a lightweight latent-space surrogate for efficient gradient computation. The method has demonstrated superior sample efficiency compared to PPO on manipulation tasks, including object manipulation with a quadruped robot. AI
IMPACT Enables training RL policies in complex, hard-to-model environments without physics simulators, potentially accelerating robotics and manipulation research.