Uranus: Building the Next-Generation Simulation Infrastructure for Embodied AI

Uranus. Building the Next-Generation Simulation Infrastructure for Embodied AI. By Wenkang Qin and seven others, from D-Robotics. Published on the twenty-third of September. Its code repository is public and real: about thirty source files.

Scalable simulation matters for generating robot data, training policies, evaluating them, and iterating safely. Real-world interaction is costly, and conventional simulators require labor-intensive construction. Something that scales has to avoid both. It cannot lean on expensive time with real hardware, and it cannot wait on people to build every scene by hand.

Uranus is a data-driven robot simulator. At its core is an autoregressive diffusion model conditioned on joint trajectories. Instead of a world built by hand, the simulator learns to generate what happens given how the robot's joints move.

The authors describe three capabilities. First, streaming, open-ended rollout. Uranus receives future joint-position trajectories online and autoregressively generates one latent frame per step, corresponding to four RGB frames, with no fixed horizon. Second, low-latency generation, reaching twenty-four frames per second after inference optimization. Third, scalable and extensible robot control, through a unified interface for synchronized multi-view generation across diverse robot embodiments and camera configurations.

Evaluation is both quantitative and qualitative, on in-distribution and out-of- distribution data. The authors present it as an objective assessment that also clearly identifies the simulator's current limitations. The code and the model weights are released.

The design swaps construction for learning. A step at a time, fed only the trajectory the joints will follow, Uranus produces the views across cameras and embodiments, and it keeps going for as long as trajectories keep arriving. That is what makes it infrastructure rather than a demonstration. A simulator driven this way can sit underneath data generation, policy training and evaluation, the same jobs the problem began with.