Nebius Physical AI Workbench brings agentic workflows to physical AI builders in collaboration with NVIDIA
Nebius Physical AI Workbench brings agentic workflows to physical AI builders in collaboration with NVIDIA
The Nebius Physical AI Workbench is bringing AI agent workflows to robotics development, with customers reporting it cuts engineering time by up to 40%. Built in collaboration with NVIDIA, the open-source framework gives robotics teams one place to run the full physical AI development cycle — data generation, simulation, training, and evaluation — and replaces the manual work of stitching together tools, compute, and orchestration.
Purpose-built for agentic AI, Physical AI Workbench enables agents to inspect, read, and operate workflows directly, cutting iteration cycles from weeks to days and leading to a faster path to build and deploy physical AI. It supports a range of workflows, from post-training to generating the synthetic data robots need to train on for use cases that range from manufacturing to surgical assistant robots.
“Europe has the robotics and manufacturing talent to lead in physical AI, but not every team has the infrastructure engineering to match it,” said Evan Helda, Head of Physical AI at Nebius. “Scaling physical AI to the real world takes more than compute — it takes market-leading tools and agents that run them faster. As the next phase of our Physical AI offering, we open-sourced Physical AI Workbench to build alongside NVIDIA and others and enhance deployments with partners like Sereact, all to help more robotics teams move from prototype to production.”
Nebius collaborates with NVIDIA and other physical AI partners
One system on Nebius gives robotics and physical AI teams the leading physical AI tools, verified and integrated by Nebius engineers, so they can build and deploy physical AI faster.
Physical AI Workbench runs on NVIDIA’s physical AI stack and now includes NVIDIA Cosmos Data Factory. Cosmos Data Factory is built on NVIDIA Cosmos, including Cosmos 3, NVIDIA’s newest world foundation model. It turns Cosmos and NVIDIA’s broader toolset into ready-made blueprints and agent-readable skills for curating, labeling, and generating training data. Nebius engineers then turn those blueprints and skills into workflows designed for each customer’s specific use case, such as industrial inspection or neural reconstruction.
Physical AI Workbench operates with Isaac Sim and Isaac Lab for simulation, Isaac GR00T for robot policy models, and NVIDIA SONIC, a humanoid behavior model trained on large-scale human motion data. Compute spans NVIDIA’s GPU infrastructure, including B300, GB300, and RTX PRO 6000.
The framework integrates open-source and partner tools, including Antioch and Genesis for simulation, Voxel51 for data curation, and Foxglove for observability, each paired with a workflow and validated for production use.
“Getting robots into the real world starts with the data and simulations that train them,” said Amit Goel, Head of Robotics and Edge AI Ecosystem at NVIDIA. “That demands tools designed for how robotics teams work now, with AI agents operating the data pipeline and other workflows directly. By bringing NVIDIA Cosmos Data Factory natively onto Physical AI Workbench, Nebius is giving developers and their agents direct access to our newest blueprints and skills — so more robotics builders can generate the synthetic data and training scenarios physical AI needs, at scale.”
Nebius and Sereact partner to accelerate enterprise robotics adoption across Europe
Nebius customers are already using Physical AI Workbench today to deploy physical AI in the real world including Roboforce, Moon Surgical, Bellboy Robotics, and Sereact. Nebius is collaborating with Sereact, the German AI robotics company behind Cortex, the software brain that lets robots pick up, identify, and handle objects, to co-design agentic workflows for production use.
Sereact already deploys robots in production for customers across retail and automotive, including Zalando, CEVA, IKEA, DHL, BMW, and Mercedes-Benz. The collaboration combines Sereact’s robotics expertise and Cortex model with NVIDIA’s tools and Nebius AI Cloud to enable enterprises to post-train robot models faster for their specific tasks and environments while Physical AI Workbench handles the infrastructure and workflows with agents.
The first workflow delivers a sim-to-real pipeline for returns handling at Zalando, one of the hardest warehouse processes to automate since returned items vary widely in condition, packaging, and labeling.
“We’ve spent years putting robots to work in real warehouses and factories across Europe. What we needed was better simulation. Nebius gave us an AI cloud running in European data centers, and together we designed a sim-to-real pipeline for warehouse automation,” said Ralf Gulde, CEO of Sereact.
Nebius Physical AI Workbench is available now on GitHub
Nebius is actively developing and deploying the open-source framework through the Physical AI Living Lab, launched with NVIDIA in June 2026, where European robotics companies build and test physical AI with direct support from Nebius and NVIDIA engineers.
Nebius Physical AI Workbench is available now on GitHub. To learn more, see Nebius physical AI solutions.
Explore Nebius Token Factory
See also
The Nebius Physical AI Workbench turns NVIDIA Cosmos 3, NVIDIA Isaac Sim, NVIDIA Isaac GR00T and other Physical AI tools into composable building blocks that agents can wire together. We are building it in the open, and it is available now on GitHub.
Nebius Academy and NVIDIA are launching an education collaboration at GTC Berlin, featuring: NVIDIA Deep Learning Institute (DLI) open-source courses hosted on Nebius Academy starting with Fundamentals of Deep Learning, co-developed courses in agentic and physical AI that run on Nebius, and Nebius GPU compute behind NVIDIA DLI’s sold-out full-day workshop at GTC.
SemiAnalysis has named Nebius a Platinum AI cloud provider, the top tier of its ClusterMAX™ 3.0 rating. The rating is based on live testing of production clusters: burn-ins, storage under concurrent load, and automated failure recovery on real NVIDIA GB300 and NVIDIA HGX B300 hardware. This post covers what the testing found and why it matters for teams choosing where to run production AI.