Repolode
Submit
nebius

nebius/nebius-physical-ai

A control plane that lets coding agents and the npa CLI run physical-AI workloads, such as simulation, training and data curation, on Nebius Kubernetes.

View on GitHub
Stars
50
Forks
19
Open issues
36
Last commit
4 minutes ago

What is it?

Nebius Physical AI is a control plane for running physical-AI workloads on Nebius Cloud. The npa CLI and Python SDK configure a project, check access, plan GPU and CPU resources, submit containerized tools and multi-stage workflows, and inspect the resulting artifacts. Workflows are defined in YAML as state graphs, where each stage calls a Workbench tool and its outputs feed later stages. SkyPilot schedules the jobs onto Nebius Kubernetes, while Object Storage holds inputs, checkpoints, media and run state.

The project covers a range of tasks, including Cosmos generation, video augmentation and curation, neural 3D reconstruction, robot-policy training with LeRobot, and reinforcement learning with Isaac Lab, Genesis and MJLab. It also offers hosted inference through Token Factory, which needs no customer GPU. Coding agents such as Codex or Claude Code can drive the same operations from a checkout of the repository.

A shared deployment mode adds personal namespaces, GPU caps and scoped storage for teams, with sign-in through npa login.

Key features

  • 01Describes workloads as YAML workflow state graphs that chain containerized tool stages.
  • 02Submits GPU and CPU jobs to Nebius Kubernetes through SkyPilot.
  • 03Stores inputs, checkpoints, media and run state in S3-compatible Object Storage.
  • 04Provides a CLI and Python SDK for configuration, preflight checks, planning and run inspection.
  • 05Supports hosted inference through Token Factory without a local GPU.
  • 06Includes teardown commands that preview project cleanup before deleting resources.

Use cases

  • Generating or augmenting robot training video with Cosmos models on cloud GPUs.
  • Training a robot policy with LeRobot from a recorded dataset.
  • Running Isaac Lab or Genesis reinforcement learning jobs on Nebius GPU nodes.
  • Letting a team share a Workbench endpoint with per-user GPU limits and scoped storage.

Topics

#physical-ai#kubernetes#skypilot#robotics#gpu#workflow#object-storage#reinforcement-learning

FAQ

What is nebius-physical-ai?
A control plane that lets coding agents and the npa CLI run physical-AI workloads, such as simulation, training and data curation, on Nebius Kubernetes. Nebius Physical AI is a control plane for running physical-AI workloads on Nebius Cloud. The npa CLI and Python SDK configure a project, check access, plan GPU and CPU resources, submit containerized tools and multi-stage workflows, and inspect the resulting artifacts. Workflows are defined in YAML as state graphs, where each stage calls a Workbench tool and its outputs feed later stages. SkyPilot schedules the jobs onto Nebius Kubernetes, while Object Storage holds inputs, checkpoints, media and run state.
Is nebius-physical-ai free to use? What license does it have?
nebius-physical-ai is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is nebius-physical-ai still maintained?
Actively maintained. The last commit to nebius-physical-ai was on Oct 22, 2026.
What can I use nebius-physical-ai for?
Generating or augmenting robot training video with Cosmos models on cloud GPUs. Training a robot policy with LeRobot from a recorded dataset. Running Isaac Lab or Genesis reinforcement learning jobs on Nebius GPU nodes. Letting a team share a Workbench endpoint with per-user GPU limits and scoped storage.
Where can I get nebius-physical-ai?
The source code and installation instructions are on GitHub: https://github.com/nebius/nebius-physical-ai

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects