Uranus: Building the Next-Generation Simulation Infrastructure for Embodied AI
Uranus is a robot simulator that generates open-ended, multi-view rollouts from planned joint trajectories at 24 FPS.
The paper frames the system as a data-driven alternative to hand-built simulation for robot training, evaluation, and safer iteration. Its core model is a joint-trajectory-conditioned autoregressive diffusion model that produces one latent frame per step, corresponding to four RGB frames. The authors say it supports varied robot embodiments and camera setups through a unified control interface, and they release code, data, SDK code, and model weights. ArXiv · AI/CL/LG's note
The paper frames the system as a data-driven alternative to hand-built simulation for robot training, evaluation, and safer iteration. Its core model is a joint-trajectory-conditioned autoregressive diffusion model that produces one latent frame per step, corresponding to four RGB frames. The authors say it supports varied robot embodiments and camera setups through a unified control interface, and they release code, data, SDK code, and model weights. ArXiv · AI/CL/LG's note
score 6