DriftWorld

DriftWorld is an action-conditioned world model for robotics introduced by researchers from MIT and Harvard University. It uses drifting generative models (building on the 'Generative Modeling via Drifting' framework) instead of iterative diffusion, learning an action-conditioned drift field during training so that future video frames can be generated from the current observation and candidate action sequence in a single forward pass. DriftWorld generates at 30+ fps on a single H100 GPU (33.3 fps on Bridge-V2, 38.8 fps on RT-1, 36.6 fps on Language Table, 100 fps on Robomimic, 270.3 fps on Push-T), achieving an average 17x speedup over diffusion-based baselines while matching or exceeding them in visual quality (SSIM, PSNR, LPIPS, FID, FVD). It achieves state-of-the-art performance on multiple robot manipulation benchmarks including Bridge-V2, RT-1, Language Table, Push-T, and Robomimic. DriftWorld can also serve as an offline simulator for ranking real-world robot policies with Pearson correlations up to 0.99 versus ground-truth performance.