A curated list of resources for NVIDIA's modular physical-simulation stack — spanning OpenUSD, Omniverse, DSX, robotics, and CAE/CFD workflows.
This list is organized around a layered infrastructure view of the stack, not as a single monolithic platform. The insight: NVIDIA's industrial significance lies in connecting interoperable scene data, GPU-native simulation, and blueprint-packaged deployment — not in any single simulator alone.
This repository is accompanied by a working paper currently in draft:
We're actively seeking co-authors and contributors. If you work on any part of this stack — at NVIDIA, at a partner company, in research, or in open-source — your perspective would strengthen the analysis. See paper/CALL_FOR_AUTHORS.md for details.
| Area | What's Needed |
|---|---|
| OpenUSD standards & AOUSD | Governance process, spec evolution, cross-vendor adoption data |
| DSX deployment | Production experiences, scaling observations, integration patterns |
| Robotics (Isaac Sim / Lab / OSMO) | Sim-to-real transfer results, benchmark comparisons, workflow reviews |
| CAE / CFD digital twins | PhysicsNeMo surrogate model accuracy, Kit-CAE production readiness |
| Newton & Warp | Differentiable simulation use cases, performance benchmarking |
| Competitor/alternative analysis | Comparison with Unity, MuJoCo, Gazebo, AWS IoT TwinMaker, Azure Digital Twins |
Physical AI, industrial digital twins, and AI-factory operations all place new pressure on simulation systems. The challenge is no longer just running a physics kernel quickly — it's organizing data exchange, scene composition, simulation services, delivery, and vertical deployment so multiple teams can work on a shared operational model.
+-----------------------------------------------------+
| Vertical Workflows DSX . Robotics . CAE/CFD |
+-----------------------------------------------------+
| Delivery & Orchestration App Streaming . Helm |
+-----------------------------------------------------+
| Runtime & App Construction Kit SDK . Libraries |
+-----------------------------------------------------+
| Acceleration & Simulation Warp.PhysX.RTX.Newton |
+-----------------------------------------------------+
| Interoperability & Semantics OpenUSD . SimReady |
+-----------------------------------------------------+
- Layer 1: Interoperability & Semantics
- Layer 2: Acceleration & Simulation
- Layer 3: Runtime & Application Construction
- Layer 4: Delivery & Orchestration
- Layer 5: Vertical Workflows & Blueprints
- Cross-Cutting Topics
- Timeline & Ecosystem Evolution
- Community Projects & Integrations
- Ecosystem Landscape
- Learning Paths
- Contributing
OpenUSD is not a file format — it's the shared scene and composition substrate that lets data and semantics move through the entire stack.
- OpenUSD Official Site — Specification, API reference, and community resources
- OpenUSD GitHub — Pixar's reference implementation
- OpenUSD Core Specification 1.0 — Ratified December 2025 by AOUSD, the formal standards base for scene composition
- USD FAQ — Covers stages, layers, and composition arcs
- How to Use OpenUSD — NVIDIA's developer-oriented guide to authoring and composition workflows
- Exchange SDK Documentation — Tools for ingest and interchange across upstream design tools (CAD, DCC, etc.)
- Connect Samples — Reference code for building OpenUSD connectors
- SimReady Overview — OpenUSD-based building blocks with metadata, semantics, sensor attributes, and physical properties
- SimReady Asset Specification — Technical specification for simulation-ready assets
- SimReady Explorer — Browse and discover SimReady assets
- AOUSD Official Site — The industry alliance driving OpenUSD standardization
- Core Spec 1.0 Blog — Context on the ratification and what it means for the ecosystem
GPU-native simulation substrate: engines that compute physics, rendering, and domain-specific solvers.
- NVIDIA Warp — Python framework for GPU simulation and spatial computing
- Warp GitHub — Source code, examples, and documentation. Differentiable kernels work with PyTorch, JAX, and Paddle
- Warp Documentation — API reference, tutorials, and differentiable simulation examples
- Warp Publications — Academic and research papers leveraging Warp
- PhysX 5 SDK — GPU-accelerated rigid body, soft body, fluid, and cloth simulation
- PhysX GitHub — Open-source SDK
- RTX Renderer Documentation — Real-time and path-traced rendering in the Omniverse stack
- Newton GitHub — GPU-accelerated physics engine built on Warp, targeting roboticists and simulation researchers. Extends Warp's warp.sim module and integrates MuJoCo Warp
- Newton Developer Page — Official NVIDIA page
- PhysicsNeMo GitHub — Open-source framework for building, training, and fine-tuning physics-ML models (CFD, structural mechanics, electromagnetics)
- Earth-2 Studio — Inference sub-module for climate and weather AI models
- CUDA-X Libraries — GPU-accelerated libraries for linear algebra, FFT, signal processing, and more
The developer-facing layer for building applications and headless services.
- Kit SDK Overview — Developer guide for building Omniverse applications
- Kit 106.0 Release Highlights — The May 2024 pivot toward developer-facing infrastructure
- Kit App Template GitHub — Official starter for custom Kit applications
- Omniverse Libraries Overview — Modular libraries including ovrtx, ovphysx, and more
- Platform Overview — How libraries, SDKs, services, and blueprints compose into the platform
- Introduction — Developer Overview — Omniverse as a development stack for 3D applications at scale
- Extensions Documentation — Discover and build Kit extensions
How simulation applications reach users — from GPU servers to browsers.
- App Streaming Documentation — Stream Kit applications via WebRTC to browsers and thin clients
- Omniverse Container Deployment — Container and Helm-based deployment for production workloads
- NVIDIA OSMO GitHub — Developer-first platform for scaling Physical AI workloads across heterogeneous compute (training GPUs, simulation clusters, edge devices) in YAML. Powers Isaac Lab, Isaac Sim, and Project GR00T at scale
NVIDIA Blueprints package the stack into reusable, domain-oriented workflows.
DSX is the strongest full-stack case: shared OpenUSD scene data, domain simulations, App Streaming, multi-role review, and blueprint packaging in one reference workflow.
- DSX Blueprint Overview — 50-acre AI-factory geometry set, thermal/electrical simulations, multi-role workflow
- DSX Blueprint GitHub — Source code, web stack, and deployment documentation
- DSX Launch Blog — Gigawatt-scale AI factory digital twins
- Vera Rubin DSX Announcement — GA release with broad industry support (March 2026)
- Isaac Sim — Open-source reference framework for robotics simulation, testing, and synthetic data generation
- Isaac Sim GitHub — Source repositories
- Isaac Lab GitHub — GPU-accelerated unified framework for robot learning (RL, imitation learning, motion planning). 6.8k+ stars
- Isaac Lab Arena — Composable, scalable system for diverse simulation environments
- OSMO — Orchestration for the full Physical AI pipeline: training, simulation, hardware testing
- Kit-CAE Guide — Combining NVIDIA technologies for CAE data (CGNS, NPZ schemas)
- Physics Digital Twin Blueprint Blog — Omniverse APIs + PhysicsNeMo + App Streaming + reference deployment
- NVIDIA Omniverse Blueprints GitHub — Open-source physical-simulation blueprints (DSX and more)
- NVIDIA AI Blueprints GitHub — Open-source AI/agentic blueprints (RAG, digital human, drug discovery, retail, and more)
- Contributions welcome — see Contributing
- Warp Differentiable Simulation — Differentiable physics for optimization and learning
- Newton — Integrates MuJoCo Warp for differentiable rigid body simulation
- Contributions welcome for research papers and benchmarks
- Neural Reconstruction in Isaac Sim — Capture-to-USD pipelines for robotics
- Contributions welcome for NeRF/3DGS to OpenUSD tools
- NVIDIA NIM — Inference microservices composable with simulation workflows
- Contributions welcome
| Date | Milestone |
|---|---|
| May 2024 | Kit 106 pivots toward developer-facing infrastructure |
| Aug 2024 | NIM Agent Blueprints launch as customizable workflows |
| Oct 2024 | "NVIDIA Blueprints" becomes the lasting umbrella term |
| Nov 2024 | CAE digital twin blueprint makes vertical packaging explicit |
| Oct 2025 | DSX blueprint introduced for gigawatt-scale AI factories |
| Dec 2025 | OpenUSD Core Specification 1.0 ratified by AOUSD |
| Mar 2026 | DSX blueprint reaches general availability |
| Mar 2026 | NVIDIA OSMO open-sourced for Physical AI orchestration |
This is where your project goes. If you're building on any layer of this stack, submit a PR!
- awesome-isaac-gym — Curated list of Isaac Gym frameworks, papers, and resources
- Awesome-Physical-Engineering-AI — AI tools for hardware engineering, CAD, simulation, and manufacturing
- awesome-robotics-libraries — Comprehensive list of robotics simulators and libraries
- Awesome-Physics-aware-Generation — Physics-aware generative modeling papers
- Your project here
- Your project here
- Your project here
- Your project here
How NVIDIA's stack relates to adjacent platforms and alternatives.
| Layer | NVIDIA Stack | Alternatives / Complements |
|---|---|---|
| Scene standard | OpenUSD | glTF, FBX, STEP |
| Physics engine | PhysX, Newton, Warp | MuJoCo, Bullet, DART, Chrono |
| Robot sim | Isaac Sim / Lab | Gazebo, MuJoCo MPC, PyBullet, SAPIEN |
| Digital twins | DSX, Kit-CAE | Azure Digital Twins, AWS IoT TwinMaker, Eclipse Ditto |
| Orchestration | OSMO | Kubernetes + custom, Airflow |
| AI surrogates | PhysicsNeMo | DeepXDE, custom PINN frameworks |
| Rendering | RTX | Vulkan, OptiX, Embree |
- Understand OpenUSD composition arcs — stages, layers, sublayers, references, payloads, variants
- Set up Kit SDK and build a first extension
- Explore SimReady assets and understand simulation-ready metadata
- Try Warp for GPU-accelerated Python simulation
- Build a custom Kit application using the App Template
- Deploy a streamed application using App Streaming
- Run the DSX Blueprint locally
- Orchestrate a multi-step Physical AI pipeline with OSMO
- Compose multi-domain simulations using OpenUSD composition arcs
- Integrate PhysicsNeMo surrogate models into a digital twin workflow
- Build and deploy a custom NVIDIA Blueprint for your vertical
- Contribute to the working paper
Contributions are welcome and encouraged! See CONTRIBUTING.md for detailed guidelines.
What we're looking for:
- Official documentation, blog posts, and tutorials
- Open-source projects building on any layer of the stack
- Real-world case studies and deployment experiences
- Learning resources — tutorials, courses, videos
- Research papers on differentiable simulation, capture-to-sim, agentic operations, or OpenUSD
- Industry integrations and connectors
- Paper co-authorship — help refine the architecture analysis
This list is released under CC0 1.0 Universal. The working paper in paper/ is under a separate collaborative license — see paper/LICENSE.
Maintained by @tsubasakong · Inspired by awesome