README.md

metadata

title: OpenRAL
emoji: 🤖
colorFrom: indigo
colorTo: indigo
sdk: static
pinned: true
short_description: The open harness for physical AI
license: apache-2.0
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/679bec571de528e679aae409/KR7hrL_kADf-oPybj7Q0C.png

OpenRAL

The open harness for physical AI. OpenRAL is the open-source Robot Agentic Layer (harness) for embodied AI. It unifies fast policies, slow reasoning, reward signals, perceptual AI and classical control into a typed, traceable, safety-first runtime — on real hardware and in simulation. One fixed harness, swappable rSkills.

---

What is OpenRAL?

A VLA alone is not an agent. OpenRAL wraps it in the loop it needs — a typed, layered runtime between a robot's motor API and a task planner:

• The harness — a slow LLM reasoner (S2) planning in typed tool-calls, fast visuomotor policies (S1, 30–200 Hz) executing action chunks, perception and world state keeping both honest, and a separate C++ deny-by-default safety kernel gating every command. Pydantic v2 contracts at every layer boundary.
• rSkills — every capability the agent has is an rSkill, not just VLAs: detectors, scene VLMs, reward monitors, MoveIt 2 / Nav2 actions and reasoner playbooks, packaged as Hub repos with weights, a typed rskill.yaml manifest and reproducible eval. Install like a model: openral rskill install OpenRAL/rskill-smolvla-franka_panda-libero_spatial-bf16. Swap any of them; the harness never changes.
• Benchmark → sim → real — the same agent runs benchmark suites, headless sim rollouts and real hardware, and every run is an OpenTelemetry trace you can replay, audit, or fold into a LeRobot dataset flywheel.

---

What's shipped

🤖 15+ robot platforms	SO-100/SO-101, Franka Panda, UR5e/UR10e, ALOHA bimanual, Galaxea A1, OpenArm v2, Unitree H1/G1, Rethink Sawyer, Fourier GR1, WidowX
🧠 30+ rSkills	SmolVLA, π0.5, GR00T N1.7, LingBot-VLA 2.0, InternVLA-N1 (VLN), MolmoAct2, xVLA, ACT, Diffusion Policy, RLDX-1, OpenVLA-OFT, RT-DETR, OmDet-Turbo, LocateAnything, Qwen3.5-4B VLM, Robometer-4B, MoveIt 2 / Nav2 actions, playbooks
🎮 Benchmark suites	LIBERO, MetaWorld, ManiSkill3, SimplerEnv, RoboCasa, RoboTwin 2.0, RLBench, gym-aloha, gym-pusht
🔬 Sim engines	MuJoCo, Isaac Sim, Genesis, SAPIEN, robosuite, CoppeliaSim
📡 Sensors & perception	RealSense D435/D435i/D415, Luxonis OAK-D Pro, USB UVC, Robotiq FT-300 · Depth Anything 3 metric depth · cuVSLAM stereo SLAM

---

Quick start

curl -fsSL https://raw.githubusercontent.com/OpenRAL/openral/master/scripts/install.sh | bash
openral doctor
openral rskill install OpenRAL/rskill-smolvla-franka_panda-libero_spatial-bf16
openral sim run scenes/benchmark/libero_spatial.yaml

---

Community & support

• 💬 Discord — questions, help, design chat
• 📧 hello@openral.com — general enquiries
• 🔒 Security disclosures: security@openral.com — never a public issue

---

License

OpenRAL is Apache-2.0. Third-party model weights keep their own upstream licenses; the loader surfaces each weight's license at install time.