OpenRAL
OpenRAL is an open-source operating layer for embodied AI: one typed, traceable, safety-first runtime that unifies fast policies, slow reasoning, and classical control for deployable robot agents.
Quick start
git clone https://github.com/OpenRAL/openral && cd openral
just bootstrap # installs uv, ROS 2, system deps
just sync # install Python workspace (wraps uv sync + repair)
uv run openral doctor # verify your environment
just test # run the test suite
What does it do?
- Load any VLA (SmolVLA, π0.5, GR00T N1.7, OpenVLA, …) on any robot (SO-100, G1, UR5e, …).
- Type-safe, layer-isolated architecture — HAL → Sensors → World State → Skill → Reasoning → Safety → Observability.
- Skill packaging on HuggingFace Hub with versioned, license-aware manifests (sigstore provenance signing is planned, not yet implemented).
- Full OpenTelemetry traces per execution.
- LeRobotDataset v3 flywheel — every execution becomes a training row.
Navigation
Get started
Tutorials
- Your first sim rollout — start here: a real policy driving a simulated robot, no GPU needed
- Create a sim environment — author a scene, robot or adapter
- Run a benchmark — reproducible numbers, and what publishing one commits you to
- Write & publish an rSkill — package a policy
- Quantize an rSkill — fit a VLA on an 8 GB card
- Deploy on a robot & open the dashboard — real hardware
Understand the system
- Architecture overview
- Repo state map — interactive per-module status canvas (open in a browser; no build step)
Plan & status
Reference