What it is
Open workflows for training and running robot world models for policy evaluation, with data prep, inference, and checkpoint tools for 1.3B and 5B models. Train, run, and adapt robot world models that generate controllable rollouts from text, first-frame images, and control videos, with included data preprocessing, latent caching, and checkpoint merge tools.
Key results
- Supports i2v and t2v inference with 10 FPS, up to 99 frames by default
- Stage-2 DMD2 reduces denoising from 20 steps to 4-6 steps
- Runs production training on single-node 8x H20 or 8x A100 with ZeRO-2/3
Shows that an open, end-to-end stack for robot world-model training and rollout generation is practical with current video diffusion backbones.