Video-generating robot simulator runs at 24 frames per second
single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC
What this means for you
Worth a look if real-world rollouts are your bottleneck — the weights are public and generation runs in real time. Nothing to plan around yet: one preprint, no scores in the abstract, no outside reproduction, no hardware details. The authors' limitations section is the first thing to check.
Uranus is a robot simulator that generates video instead of computing physics. It is an autoregressive diffusion model — each frame is produced by denoising noise, and each step feeds on the last — conditioned on the joint-position trajectory the robot is meant to follow, with no fixed horizon. It streams at 24 frames per second after optimisation, emits four RGB frames per step, and gives one interface for synchronised multi-view generation across different robot bodies and camera layouts. Code and weights are released; the paper reports in-distribution and out-of-distribution tests without publishing figures in the abstract.
Key facts
- ·Uranus generates one latent frame per step, corresponding to four RGB frames, with no fixed rollout horizon. source
- ·It reaches 24 frames per second after inference optimisation. source
- ·The authors released both code and model weights. source
- ·Evaluations cover in-distribution and out-of-distribution data, and the paper states current limitations. source
- ·The paper carries a publication date of 22 September 2026. source
What the sources say
- Hugging Face Daily Papers (research) — Presents a learned simulator for robot data generation, with open weights and a stated limitations section.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersUranus: Building the Next-Generation Simulation Infrastructure for Embodied AI2026-09-22