Nvidia releases DreamDojo, an open-source world model for robot training

Nvidia's AI research team released DreamDojo, an interactive world model that generates simulated futures from robot motor controls without needing a physics engine or hand-authored dynamics. Nvidia's AI Director called it "Simulation 2.0." The model is pre-trained on 44,000 hours of first-person human video with zero robot-in-the-loop, using "latent actions" that capture state changes without hardware knowledge. A real-time version runs at 10fps, stable for over a minute, and supports live VR teleoperation inside the simulation.

View full digest for February 23, 2026