Uranus: Building the Next-Generation Simulation Infrastructure for Embodied AI
Abstract
Scalable simulation is essential for robot data generation, policy training, evaluation, and safe iteration, yet real-world interaction is costly and conventional simulators require labor-intensive construction. We present Uranus, a data-driven robot simulator built around a joint-trajectory-conditioned autoregressive diffusion model. Uranus offers three key capabilities: (1) streaming, open-ended rollout, which receives future joint-position trajectories online and autoregressively generates one latent frame per step, corresponding to four RGB frames, without a fixed horizon; (2) low-latency generation, achieving 24 FPS after inference optimization; and (3) scalable, extensible robot control, providing a unified interface for synchronized multi-view generation across diverse robot embodiments and camera configurations. We conduct comprehensive quantitative and qualitative evaluations on both in-distribution and out-of-distribution data, providing an objective assessment of Uranus and clearly identifying its current limitations. We release the code and model weights to empower the community with practical tools and insights.
- Comment
- Project Page: https://d-robotics-ai-lab.github.io/large-model-team/blog/uranus/ Inference Code: https://github.com/D-Robotics-AI-Lab/Uranus-OSS Inference Data: https://huggingface.co/datasets/D-Robotics/Uranus-Demo-Data SDK Code: https://github.com/D-Robotics-AI-Lab/Uranus-SDK Model Weights: https://huggingface.co/collections/D-Robotics/uranus