Chunk-wise Model-based Policy Optimization

This research proposes a model-based reinforcement learning approach that uses a chunk world model as a differentiable training-time evaluator to optimize the chunk actor without recurrent model invocation. The method generates H-step hypothetical rollouts in a single forward pass and directly optimizes value expansion, avoiding costly and unstable BPTT. It is evaluated on 19 challenging high-dimensional humanoid tasks from HumanoidBench and deployed for real-world quadruped and bipedal walking.