A robot that plans by imagining outcomes is only as useful as the number of futures it can imagine per second. Diffusion-based world models imagine beautifully but slowly, since every rollout needs many denoising steps, which throttles how many candidate action sequences you can search at decision time.

DriftWorld swaps iterative denoising for a learned drift. During training it learns how the scene should move given an action, then at inference produces future frames in a single forward pass, hitting 30-plus fps and averaging 17 times faster than diffusion baselines. Across manipulation benchmarks including Bridge-V2, RT-1, and Push-T, the speedup arrives without sacrificing decision quality. It also doubles as an offline simulator for ranking real robot policies, with rollout scores tracking ground truth up to 0.99.

Those figures are the authors' own on standard benchmarks, so the paper is worth reading for the drift formulation and its limits.