D-Robotics Debuts Uranus, a Diffusion-Based Robot Simulator

The paper opens on Uranus as "a data-driven robot simulator" built around a "joint-trajectory-conditioned autoregressive diffusion model." The Hugging Face daily list surfaced it on September 24. Instead of integrating torques or velocities, the model generates the visual consequences of a specified kinematic trajectory.

Three claims carry the pitch. The simulator supports streaming rollout that "generates one latent frame per step, corresponding to four RGB frames," it hits "24 FPS after inference optimization," and it exposes "a unified interface for synchronized multi-view generation across diverse robot embodiments and camera configurations."

The task-OOD protocol is the piece for anyone worried that embodied AI benchmarks quietly train on their test set. The authors (Wenkang Qin, Yukun Zhou, Noah Shen, Jisong Cai, Dongxiao Mao, Baicheng Li, Yue Zhang and Wei Sui) split 30 tasks into 24 for training and 6 for held-out evaluation, with a designated test case, and report "comprehensive quantitative and qualitative evaluations on both in-distribution and out-of-distribution data."

The paper page publishes no headline win-rate against a physics-engine baseline. That is the number a robot-learning team would want.

The upload landed on a busy robotics day for us. Three of today's papers touch the same stack, including MemBodied and FLUX 3 Action.