NVIDIA/OSMO

NVIDIA/OSMO

Description: The developer-first platform for scaling complex Physical AI workloads across heterogeneous compute—unifying training GPUs, simulation clusters, and edge devices in a simple YAML

Language: TypeScript

License: Apache-2.0

Stars: 174

Forks: 38

Open issues: 107

Created: 2025-10-01T18:18:06Z

Pushed: 2026-06-11T02:28:03Z

Default branch: main

Fork: no

Archived: no

README:

Welcome to OSMO

Workflow Orchestration Purpose-built for Physical AI

Get Started | Documentation | Community | Roadmap

Use OSMO to manage your workflows, version your datasets and even remotely develop on a backend node. Using OSMO's backend configuration, run your workflows seamlessly on any cloud environment. Build a data factory to manage your synthetic and real robot data, train neural networks with experiment tracking, train robot policies with reinforcement learning, evaluate your models and publish the results, test the robot in simulation with software or hardware in loop (HIL) and automate your workflows on any CI/CD systems

For Robotics & AI Developers

Write once, run anywhere. Focus on building robots, not managing infrastructure.

# Your entire physical AI pipeline in a YAML file
workflow:
tasks:
- name: simulation
image: nvcr.io/nvidia/isaac-sim
platform: rtx-pro-6000 # Runs on NVIDIA RTX PRO 6000 GPUs

- name: train-policy
image: nvcr.io/nvidia/pytorch
platform: gb200 # Runs on NVIDIA GB200 GPUs
resources:
gpu: 8
inputs: # Feed the output of simulation task into training
- task: simulation

- name: evaluate-thor
image: my-ros-app
platform: jetson-agx-thor # Runs on NVIDIA Jetson AGX Thor
inputs:
- task: train-policy # Feed the output of the training task into eval
outputs:
- dataset:
name: thor-benchmark # Save the output benchmark into a dataset

• ✅ Zero-Code Workflows – Write workflows in YAML and iterate, not Python scripts
• ✅ Truly Portable – Same workflow runs on laptop (Docker/KIND) or cloud (EKS/AKS/GKE)
• ✅ Interactive Development – Launch VSCode, Jupyter, or SSH & develop remotely on cloud
• ✅ Smart Storage – Content-addressable datasets with deduplication save 10-100x on storage
• ✅ Infrastructure-Agnostic – Workflows never reference specific infrastructure—scale transparently

For Platform & Infrastructure Engineers

Scale infrastructure independently. Add compute backends without disrupting developers.

• ✅ Centralized Control Plane – Single pane of glass for heterogeneous compute across clouds and regions
• ✅ Plug-and-Play Backends – Register new Kubernetes clusters dynamically via CLI
• ✅ Geographic Distribution – Deploy compute wherever it's available—cloud, on-prem, edge
• ✅ Zero-Downtime Changes – Scale GPU compute clusters without affecting users or their workflows

Solving Physical AI

Physical AI development uniquely requires orchestrating three types of compute working together:

| 🧠 Training | 🌐 Simulation | 🤖 Edge | |:---:|:---:|:---:| | GB200, H100 | L40, RTX Pro | Jetson AGX Thor | | Deep learning & RL | Physics & Sensor Rendering | Hardware-in-the-Loop | | Cloud | Cloud | On Premise |

Traditionally, orchestrating workflows across these heterogeneous systems requires custom scripts, infrastructure expertise, and separate tooling for each environment.

OSMO solves this Three Computer Problem for robotics by orchestrating your entire Physical AI pipeline — from training to simulation to hardware testing all in a simple YAML. No custom scripts, no infrastructure expertise required. OSMO orchestrates tasks across heterogeneous Kubernetes clusters, managing dependencies and resource allocation. By solving this fundamental problem, OSMO brings us one step closer towards making Physical AI a reality.

Key Benefits

| What You Can Do | Example | |---------------------|----------------------| | Interactively develop on remote GPU nodes with VSCode, SSH, or Jupyter notebooks | Interactive Workflows | | Generate synthetic data at scale using Isaac Sim or custom simulation environments | Isaac Sim SDG | | Train models with diverse datasets across distributed GPU clusters | Model Training | | Train policies for robots using data-parallel reinforcement learning | Reinforcement Learning | | Validate models in simulation with hardware-in-the-loop testing | Hardware In The Loop | | Transform and post-process data for iterative improvement | Working with Data | | Benchmark system software on actual robot hardware (NVIDIA Jetson, custom platforms) | Hardware Testing |

Battle-Tested in Production

OSMO is production-grade and proven at scale. Originally developed to power Physical AI workloads at NVIDIA—including Project GR00T, Isaac Lab, Isaac Dexterity, Isaac Sim, and Isaac ROS—it orchestrates thousands of GPU-hours daily across heterogeneous compute spanning cloud training clusters to edge devices.

Now open-source and ready for your robotics workflows. Whether you're building humanoid robots, autonomous vehicles, or warehouse automation systems, OSMO provides the same enterprise-grade orchestration used in production at scale.

Ready to Begin?

Select one of the deployment options below depending on your needs and environment to get started

Deploying on Microsoft Azure? Get started with [Azure NVIDIA Reference…