World models stop stepping one at a time: one forward pass yields H-step hypothetical rollouts for humanoid and quadruped control

Model-based reinforcement learning no longer has to invoke the model recurrently step by step; a chunk world model serves as a differentiable training-time evaluator that optimizes a chunk actor. Earlier paths either used derivative-free planning and imagined rollouts or dynamics gradient-based policy updates, and all of them suffer to varying degrees from recurrent compounding errors, high inference-time latency, and instability in backpropagation through time. The method is evaluated on 19 challenging high-dimensional humanoid tasks from HumanoidBench and deployed to the real world for quadruped and bipedal walking.

The world model and policy predict collaboratively in chunks, so H-step hypothetical rollouts can be generated in a single forward pass. The chunk world model is a differentiable training-time evaluator, and the chunk actor is optimized by it without recurrent model invocation. Evaluation covers 19 high-dimensional humanoid tasks from HumanoidBench, plus a real-world deployment for quadruped and bipedal walking.

This report covers what the summary states: the method is evaluated on 19 high-dimensional humanoid tasks from HumanoidBench and deployed to the real world for quadruped and bipedal walking. It suits researchers and engineers working on model-based reinforcement learning and continuous-control robotics. The summary states no next-step experiment plan.

The summary says earlier paths suffer to varying degrees from recurrent compounding errors, inference-time latency, and BPTT instability, but gives no specific comparison numbers across methods; how chunk-wise prediction behaves on long-horizon tasks is also not stated in the summary.