NVIDIA drops open-source DreamDojo robot world model

DreamDojo attacks the simulation-speed bottleneck in robotic planning by employing a clever distillation approach for real-time generation. Pre-training on 44,000 hours of human video provides a massive prior for complex physical interactions without requiring expensive, real-world robot data gathering. Reaching 10 FPS generation enables practical, real-time long-horizon planning, which has been a significant hurdle for previous, slower world models. Fully open-sourcing the model and pipeline will accelerate the broader robotics community's shift toward scalable foundation models.