World-model planning drops CEM: sampling tree search is 13.8x faster in simulation and lifts real-robot block pushing by 20%

World-model planning gains a sampling-based tree-search route that uses the learned model as the propagator of a tree search in latent space, instead of handing the problem to optimizers such as CEM and MPPI. Planning with world models has typically been done by optimizers such as CEM and MPPI, which stall on long-horizon tasks where progress requires motion the objective does not reward, such as moving around a block before pushing it. Across four simulated domains and a real-world block-pushing task, evaluated with an off-the-shelf DINO-WM, the sampling-based approaches match or exceed the success rate of CEM and MPPI.

The approach moves sampling strategies devised for high-dimensional search in configuration space directly into the latent space of a world model, so the search expands rapidly through intermediate states. Those intermediate states are ones myopic optimizers never sample, and the sampling strategies were originally devised for high-dimensional search in configuration space. Two latent-space exploration strategies are presented, nearest-neighbor-guided RRT and projection-based KPIECE, with GPU-native batched expansion for efficient device utilization.

This report covers summary-level material; the evaluation setting is four simulated domains and a real-world block-pushing task, with an off-the-shelf DINO-WM as the world model. It suits researchers and engineers in embodied planning, robot manipulation, and model-based planning who want to judge whether sampling-based tree search belongs in their own world-model planning pipeline. The material states no next-step experiment.

The material is summary-level and gives no per-domain success rates, absolute planning times, or real-robot trial counts. The real-robot result is reported only on the block-pushing task, and the 20% gain is a success-rate change relative to CEM and MPPI. The 13.8x figure is the best planning-time ratio in simulation, not a uniform factor across domains.