Nebius 发布 Physical AI Workbench：开源机器人开发框架
Nebius 与 NVIDIA 合作推出开源框架，用单一 CLI 串联数据生成、仿真、训练与评估，编码 agent 可直接操作，客户称工程时间减少最多 40%。
JUST IN: @nebiusai just launched Physical AI Workbench, an open-source framework built in collaboration with NVIDIA that gives robotics teams one place to run the entire physical AI development cycle. Data generation. Simulation. Training. Evaluation. All connected. All agent-operable. The problem it solves is real. Building physical AI means stitching together tools that each have their own dependencies, data formats and infrastructure requirements. Moving outputs between stages takes custom scripts. Repeating an experiment requires careful tracking of every scene, dataset and checkpoint. When one tool changes, the whole pipeline breaks. Workbench replaces all of that with a single CLI — npa — that a coding agent can operate directly. The results customers are already reporting: → Up to 40% reduction in engineering time → Iteration cycles cut from weeks to days → 14-stage sim-to-real reference workflow, from task definition to a checkpoint ready for hardware testing The NVIDIA integration is deep. Workbench now includes NVIDIA Cosmos Data Factory built on Cosmos 3, runs on Isaac Sim and Isaac Lab for simulation, Isaac GR00T for robot policy models, and NVIDIA SONIC for humanoid behavior. The first real-world proof point is with Sereact, the German AI robotics company already deployed with Zalando, IKEA, BMW, DHL and Mercedes-Benz. Together they built a sim-to-real pipeline for returns handling at Zalando, one of the hardest warehouse processes to automate given the variability of returned items. Available now on GitHub under Apache 2.0. The infrastructure layer of physical AI is being built in Europe. 🇪🇺📊 Learn more here: https://nebius.com/blog/posts/physical-ai-workbench ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → http://ziegler.substack.com