Open-source Physical AI stack map: 49 projects by layer
Simulation, robotics middleware, data and training, model serving, orchestration and observability, and policy and governance: the open-source projects that fill each layer, with their license and links.
This map lists 49 open-source projects across six layers of a Physical AI stack: simulation, robotics middleware, data and training, model serving, orchestration and observability, and policy and governance. Each entry gives its role, license and official links. Every project was checked on October 5, 2026: official repository found, not archived, and a commit in the last 90 days.
A Physical AI stack has many parts: a simulator, the software that runs on the robot, data and training pipelines, model serving, the infrastructure that schedules and watches all of it, and the controls that decide who can reach what. Most of those parts have mature open-source projects. This page maps them by layer, with each project's role, license and official links, so you can see the options and their terms in one place.
What does an open-source Physical AI stack contain?
Six layers, from the virtual world to the controls around it. The tables below list the projects in each.
| Layer | Projects |
|---|---|
| Simulation and physics | MuJoCo, Gazebo, NVIDIA Isaac Sim, NVIDIA Isaac Lab, Newton, Genesis, ManiSkill, Webots, Drake |
| Robotics middleware and control | ROS 2, Navigation2 (Nav2), MoveIt 2, Eclipse Zenoh, Eclipse Cyclone DDS |
| Data and training | LeRobot, openpi, NVIDIA Isaac GR00T, PyTorch, Ray, Rerun, MCAP, Apache Iceberg, Trino |
| Model serving | KServe, vLLM, NVIDIA Triton Inference Server, ONNX Runtime, Qdrant, Milvus |
| Orchestration and observability | Kubernetes, NVIDIA OSMO, Argo Workflows, Argo CD, Kueue, KubeRay, Prometheus, OpenTelemetry, Grafana, NVIDIA DCGM-Exporter |
| Policy and governance | Open Policy Agent, OPA Gatekeeper, Kyverno, Keycloak, OAuth2 Proxy, SPIFFE and SPIRE, OpenBao, cert-manager, OpenLineage, Sigstore Cosign |
Which open-source projects simulate robots and physics?
Simulators let teams test robots and policies in virtual worlds before they touch hardware, and generate training experience at scale. The projects differ in focus and in what they assume about hardware and tools, so check each project's documentation for what it needs.
| Project | What it does | License | Source code |
|---|---|---|---|
| MuJoCo | General-purpose physics simulator (Multi-Joint dynamics with Contact). | Apache-2.0 | google-deepmind/mujoco |
| Gazebo | Open-source robotics simulator; this repository is Gazebo Sim, the latest version. | Apache-2.0 | gazebosim/gz-sim |
| NVIDIA Isaac Sim | Open-source application on NVIDIA Omniverse for developing, simulating and testing AI-driven robots. | Apache-2.0 for the source. Building or using it needs additional NVIDIA-licensed components, such as the Omniverse Kit SDK. | isaac-sim/IsaacSim |
| NVIDIA Isaac Lab | Unified framework for robot learning with multi-physics and multi-renderer support. | BSD-3-Clause | isaac-sim/IsaacLab |
| Newton | GPU-accelerated physics engine built on NVIDIA Warp, for roboticists and simulation researchers. A Linux Foundation project. | Apache-2.0 (code) | newton-physics/newton |
| Genesis | Simulation platform for general-purpose robotics and embodied AI learning. | Apache-2.0 | Genesis-Embodied-AI/genesis-world |
| ManiSkill | GPU-parallelized robotics simulator and benchmark for manipulation skills. | Apache-2.0 | mani-skill/ManiSkill |
| Webots | Open-source robot simulator. | Apache-2.0 | cyberbotics/webots |
| Drake | Model-based design and verification toolbox for robotics. | BSD-3-Clause | RobotLocomotion/drake |
Which open-source robotics middleware should you know?
Middleware moves messages between the software components of a robot, and frameworks on top of it handle navigation and manipulation. ROS 2 is the common base. Zenoh and Cyclone DDS appear here as communication layers that ROS 2 can use.
| Project | What it does | License | Source code |
|---|---|---|---|
| ROS 2 | Libraries and tools for building robot applications, from drivers to algorithms and developer tools. | Apache-2.0 for the core libraries (for example rclcpp). Each package declares its own license. | ros2/ros2 |
| Navigation2 (Nav2) | Navigation framework and system for ROS 2. | Per package: Apache-2.0, BSD-3-Clause or LGPL-2.1-or-later | ros-navigation/navigation2 |
| MoveIt 2 | Motion planning, manipulation and kinematics framework for ROS 2. | BSD-3-Clause | moveit/moveit2 |
| Eclipse Zenoh | Unifies data in motion, in use and at rest. Also available as a ROS 2 middleware (rmw_zenoh). | Dual-licensed: Apache-2.0 and EPL-2.0 | eclipse-zenoh/zenoh |
| Eclipse Cyclone DDS | Open-source implementation of the OMG DDS specification, and a tier-1 middleware for ROS 2. | EPL-2.0 or EDL-1.0 | eclipse-cyclonedds/cyclonedds |
Which open-source projects handle robot data and training?
Physical AI teams collect sensor logs, curate datasets and train policies. These projects cover robot-learning libraries and open robot models, general machine-learning and distributed-compute frameworks, formats and tools for robot data, and table and query layers for large datasets.
| Project | What it does | License | Source code |
|---|---|---|---|
| LeRobot | Library for end-to-end robot learning, making AI for robotics more accessible. | Apache-2.0 | huggingface/lerobot |
| openpi | Open-source robotics models (π₀, π₀-FAST, π₀.₅) and packages from Physical Intelligence. | Apache-2.0 (repository) | Physical-Intelligence/openpi |
| NVIDIA Isaac GR00T | Foundation model for generalist robots. | Code: Apache-2.0. Model weights: NVIDIA Open Model License. | NVIDIA/Isaac-GR00T |
| PyTorch | Tensors and dynamic neural networks in Python with strong GPU acceleration. | BSD-style (see its LICENSE file) | pytorch/pytorch |
| Ray | AI compute engine: a distributed runtime and libraries for accelerating ML workloads. | Apache-2.0 | ray-project/ray |
| Rerun | Visualize, query and stream multimodal robotics data for training. | Dual-licensed: MIT and Apache-2.0 | rerun-io/rerun |
| MCAP | Modular, serialization-agnostic container file format for pub/sub and robotics data. | MIT | foxglove/mcap |
| Apache Iceberg | Open table format for huge analytic datasets. | Apache-2.0 | apache/iceberg |
| Trino | Distributed SQL query engine for big data. | Apache-2.0 | trinodb/trino |
Which open-source projects serve models?
Trained models have to run somewhere: in a cluster beside the rest of the stack, or on the robot. These projects serve or run models, and search the vectors that retrieval depends on.
| Project | What it does | License | Source code |
|---|---|---|---|
| KServe | Inference platform for generative and predictive AI on Kubernetes. | Apache-2.0 | kserve/kserve |
| vLLM | High-throughput, memory-efficient inference and serving engine for LLMs. | Apache-2.0 | vllm-project/vllm |
| NVIDIA Triton Inference Server | Inference server for cloud and edge deployments. | BSD-3-Clause | triton-inference-server/server |
| ONNX Runtime | Cross-platform accelerator for ML inferencing and training. | MIT | microsoft/onnxruntime |
| Qdrant | Vector database and vector search engine. | Apache-2.0 | qdrant/qdrant |
| Milvus | Cloud-native vector database for scalable approximate nearest-neighbor search. | Apache-2.0 | milvus-io/milvus |
Which open-source projects orchestrate and observe the stack?
Orchestration decides what runs where and in what order. Observability shows what happened. Kubernetes is the common foundation; the other projects here run, deploy, queue and monitor work on it.
| Project | What it does | License | Source code |
|---|---|---|---|
| Kubernetes | Container scheduling and management. | Apache-2.0 | kubernetes/kubernetes |
| NVIDIA OSMO | Orchestrator for Physical AI workflows: data generation, training, simulation, evaluation and hardware-in-the-loop testing. | Apache-2.0 | NVIDIA/OSMO |
| Argo Workflows | Workflow engine for Kubernetes. | Apache-2.0 | argoproj/argo-workflows |
| Argo CD | Declarative continuous deployment for Kubernetes. | Apache-2.0 | argoproj/argo-cd |
| Kueue | Kubernetes-native job queueing and scheduling. | Apache-2.0 | kubernetes-sigs/kueue |
| KubeRay | Toolkit to run Ray applications on Kubernetes. | Apache-2.0 | ray-project/kuberay |
| Prometheus | Monitoring system and time series database. | Apache-2.0 | prometheus/prometheus |
| OpenTelemetry | Open standard and tools for emitting and collecting telemetry such as traces, metrics and logs. This repository is the Collector. | Apache-2.0 | open-telemetry/opentelemetry-collector |
| Grafana | Observability and data visualization platform. | AGPL-3.0 | grafana/grafana |
| NVIDIA DCGM-Exporter | NVIDIA GPU metrics exporter for Prometheus, using DCGM. | Apache-2.0 | NVIDIA/dcgm-exporter |
Which open-source projects cover policy and governance?
Policy and governance decide who and what may reach data, models and secrets, and keep a record of it. These projects cover policy engines, identity, secrets, certificates, signing and data lineage.
| Project | What it does | License | Source code |
|---|---|---|---|
| Open Policy Agent | General-purpose policy engine. | Apache-2.0 | open-policy-agent/opa |
| OPA Gatekeeper | Policy controller for Kubernetes. | Apache-2.0 | open-policy-agent/gatekeeper |
| Kyverno | Policy as code, with Kubernetes-native policies. | Apache-2.0 | kyverno/kyverno |
| Keycloak | Identity and access management for applications and services. | Apache-2.0 | keycloak/keycloak |
| OAuth2 Proxy | Reverse proxy that provides authentication with OpenID Connect and other identity providers. | MIT | oauth2-proxy/oauth2-proxy |
| SPIFFE and SPIRE | Cryptographic identities for workloads (SPIFFE), with SPIRE as its runtime environment. | Apache-2.0 | spiffe/spire |
| OpenBao | Secrets manager: a fork of Vault managed by the Linux Foundation's OpenSSF. | MPL-2.0 | openbao/openbao |
| cert-manager | Automatic provisioning and management of TLS certificates in Kubernetes. | Apache-2.0 | cert-manager/cert-manager |
| OpenLineage | Open standard for lineage metadata collection. | Apache-2.0 | OpenLineage/OpenLineage |
| Sigstore Cosign | Code signing and transparency for containers and binaries. | Apache-2.0 | sigstore/cosign |
How did we choose these projects?
A project is on this map when it meets all four tests, checked on the date below:
- Open source. Its source code is public under a license stated in its repository. The license column gives that license in the project's own wording; where a project has several, or ships parts under other terms, the column says so.
- Official. We link the project's own repository and site, not a mirror or someone else's fork of it.
- Active. The repository is not archived, and its default branch has a commit in the 90 days before the check date (July 7, 2026 or later).
- A job in the stack. It does one of the jobs in the six layers. General-purpose infrastructure such as Kubernetes, PyTorch and Prometheus is included because a Physical AI stack needs that job done.
The order inside a layer is not a ranking, and the list is not complete: it is the projects that met these tests, not every project that exists. Permissive licenses (Apache-2.0, BSD-3-Clause, MIT) dominate. A few are copyleft: Grafana is AGPL-3.0, some Navigation2 packages are LGPL-2.1-or-later, and the licenses of Zenoh, Cyclone DDS and OpenBao (EPL-2.0, MPL-2.0) are weak copyleft.
Left out at this check. Three projects we looked at had no commit to their default branch in the 90 days before the check date: Bullet Physics (bulletphysics/bullet3, last commit October 22, 2025), OpenVLA (openvla/openvla, March 23, 2025) and Octo (octo-models/octo, July 31, 2024). That is a statement about commit dates, not about whether the code still works.
When was this list last checked?
Last checked: October 5, 2026. For each project we confirmed that the official repository exists and is not archived, found the date of its latest commit on the default branch, read its license file and README, and opened its official site. Each source below carries the same check date. Projects change, so treat the list as a snapshot and check a project's own pages before you depend on it.
Where does Hyperlake fit?
Hyperlake does not replace these projects. It is a sovereign AI operating platform that assembles, deploys and governs the data, models, applications and tools behind Physical AI, in a customer's environment or a client's, on a Kubernetes foundation. Its own pages name some of the projects above as parts of the stacks it deploys (Open Policy Agent, Apache Iceberg, Trino, Qdrant and Milvus) and say it integrates tools such as NVIDIA OSMO, Omniverse, Isaac Sim and Isaac Lab, subject to licensing and validated integration. Listing a project here is not a statement that Hyperlake supports it. For the bigger picture, see the guide to Physical AI infrastructure.
See it live. A live demonstration: real deployed models analyze recorded robot feeds, with every result tied to evidence. Open the Physical AI reliability and safety operations demonstration.
How do you suggest a project?
Know a project that meets the four tests and is missing, or an entry that is out of date? Suggest a project by email, or use the contact form. Name the project, link its official repository and say which layer it fits.
Frequently asked questions
Which open-source projects can I use to build a Physical AI stack?
Each layer has several. For simulation, MuJoCo, Gazebo, Isaac Sim, Newton and Genesis. For middleware, ROS 2 with Nav2 and MoveIt 2. For data and training, LeRobot, PyTorch and Ray. For serving, KServe, vLLM and Triton. For orchestration and observability, Kubernetes, NVIDIA OSMO, Argo and Prometheus. For policy and governance, Open Policy Agent, Keycloak and SPIFFE/SPIRE. The tables on this page list all 49 with their licenses and links.
How were the projects chosen and verified?
A project is listed if its source is public under a stated license, its official repository is not archived, and its default branch has a commit in the 90 days before the check date, and it fills a job in one of the six layers. We opened each official repository and site on October 5, 2026 and read its license file or README. Projects that failed the activity test are named on this page.
Can I use these projects commercially?
Most use permissive licenses (Apache-2.0, BSD-3-Clause, MIT), but not all. Grafana is AGPL-3.0, some Navigation2 packages are LGPL-2.1-or-later, and Zenoh, Cyclone DDS and OpenBao use weak-copyleft licenses (EPL-2.0, MPL-2.0). Some projects also ship parts under other terms: Isaac Sim needs additional NVIDIA-licensed components, and Isaac GR00T's model weights use the NVIDIA Open Model License. This page is not legal advice; read the license for your case.
Does Hyperlake build or support these projects?
No. Hyperlake does not make any of them. It is a sovereign AI operating platform that assembles, deploys and governs stacks like this in a customer's environment or a client's, and it integrates tools such as NVIDIA OSMO, Omniverse, Isaac Sim and Isaac Lab, subject to licensing and validated integration. A listing here is not a statement that Hyperlake supports a project.
How current is this list?
It was last checked on October 5, 2026, and that date is shown on this page and beside every source. Licenses, maintenance and project sites change, so check a project's own pages before you depend on it.
How do I suggest a project or report a mistake?
Use the suggestion link on this page to email the Hyperlake team, or send a message through the contact form. Name the project, link its official repository and say which layer it fits.
Sources
Sources for the facts on this page, last checked October 5, 2026.
- MuJoCo: google-deepmind/mujoco on GitHub checked October 5, 2026
- Gazebo: gazebosim/gz-sim on GitHub checked October 5, 2026
- NVIDIA Isaac Sim: isaac-sim/IsaacSim on GitHub checked October 5, 2026
- NVIDIA Isaac Lab: isaac-sim/IsaacLab on GitHub checked October 5, 2026
- Newton: newton-physics/newton on GitHub checked October 5, 2026
- Genesis: Genesis-Embodied-AI/genesis-world on GitHub checked October 5, 2026
- ManiSkill: mani-skill/ManiSkill on GitHub checked October 5, 2026
- Webots: cyberbotics/webots on GitHub checked October 5, 2026
- Drake: RobotLocomotion/drake on GitHub checked October 5, 2026
- ROS 2: ros2/ros2 on GitHub checked October 5, 2026
- Navigation2 (Nav2): ros-navigation/navigation2 on GitHub checked October 5, 2026
- MoveIt 2: moveit/moveit2 on GitHub checked October 5, 2026
- Eclipse Zenoh: eclipse-zenoh/zenoh on GitHub checked October 5, 2026
- Eclipse Cyclone DDS: eclipse-cyclonedds/cyclonedds on GitHub checked October 5, 2026
- LeRobot: huggingface/lerobot on GitHub checked October 5, 2026
- openpi: Physical-Intelligence/openpi on GitHub checked October 5, 2026
- NVIDIA Isaac GR00T: NVIDIA/Isaac-GR00T on GitHub checked October 5, 2026
- PyTorch: pytorch/pytorch on GitHub checked October 5, 2026
- Ray: ray-project/ray on GitHub checked October 5, 2026
- Rerun: rerun-io/rerun on GitHub checked October 5, 2026
- MCAP: foxglove/mcap on GitHub checked October 5, 2026
- Apache Iceberg: apache/iceberg on GitHub checked October 5, 2026
- Trino: trinodb/trino on GitHub checked October 5, 2026
- KServe: kserve/kserve on GitHub checked October 5, 2026
- vLLM: vllm-project/vllm on GitHub checked October 5, 2026
- NVIDIA Triton Inference Server: triton-inference-server/server on GitHub checked October 5, 2026
- ONNX Runtime: microsoft/onnxruntime on GitHub checked October 5, 2026
- Qdrant: qdrant/qdrant on GitHub checked October 5, 2026
- Milvus: milvus-io/milvus on GitHub checked October 5, 2026
- Kubernetes: kubernetes/kubernetes on GitHub checked October 5, 2026
- NVIDIA OSMO: NVIDIA/OSMO on GitHub checked October 5, 2026
- Argo Workflows: argoproj/argo-workflows on GitHub checked October 5, 2026
- Argo CD: argoproj/argo-cd on GitHub checked October 5, 2026
- Kueue: kubernetes-sigs/kueue on GitHub checked October 5, 2026
- KubeRay: ray-project/kuberay on GitHub checked October 5, 2026
- Prometheus: prometheus/prometheus on GitHub checked October 5, 2026
- OpenTelemetry: open-telemetry/opentelemetry-collector on GitHub checked October 5, 2026
- Grafana: grafana/grafana on GitHub checked October 5, 2026
- NVIDIA DCGM-Exporter: NVIDIA/dcgm-exporter on GitHub checked October 5, 2026
- Open Policy Agent: open-policy-agent/opa on GitHub checked October 5, 2026
- OPA Gatekeeper: open-policy-agent/gatekeeper on GitHub checked October 5, 2026
- Kyverno: kyverno/kyverno on GitHub checked October 5, 2026
- Keycloak: keycloak/keycloak on GitHub checked October 5, 2026
- OAuth2 Proxy: oauth2-proxy/oauth2-proxy on GitHub checked October 5, 2026
- SPIFFE and SPIRE: spiffe/spire on GitHub checked October 5, 2026
- OpenBao: openbao/openbao on GitHub checked October 5, 2026
- cert-manager: cert-manager/cert-manager on GitHub checked October 5, 2026
- OpenLineage: OpenLineage/OpenLineage on GitHub checked October 5, 2026
- Sigstore Cosign: sigstore/cosign on GitHub checked October 5, 2026
Start with a workload. Build the environment around it.
Tell us what you need to deploy, whose environment it must run in, and what it needs to connect to.