We've been building OpenRAL — the open-source Robot Agentic Layer (RAL), the harness for physical AI. It's the typed, traceable, safety-first runtime that holds an embodied-AI stack together: fast policies, slow reasoning, reward signals, perception, and classical control — one contract over many robots, many models, and one safety boundary, on real hardware and in simulation.
The problem: every robot SDK, VLA, and sensor speaks its own dialect, so integration is one-off glue code. Safety usually runs inside the same process that can crash. Swapping a model is a rewrite. And runs aren't reproducible.
What the harness gives you:
• One typed contract between every layer (Pydantic v2 schemas on ROS 2, tf2, MoveIt 2, Nav2, ros2_control).
• rSkills — robot skills packaged like models (HF Hub repo: typed manifest, weights, reproducible eval). Kinds: VLA policies, perception detectors, scene VLMs, reward models, ROS actions. Hot-swappable.
• Fast + slow control — slow LLM reasoner dispatches skills via typed tool-calls; fast VLA policies at 30–200 Hz.
• Perception into a live world state — open-vocab detectors + scene VLM lift 2D→3D into a tf2-aware world state + spatial-memory scene graph.
• Sim → benchmark → real, one contract — LIBERO, MetaWorld, ManiSkill3, SimplerEnv, RoboCasa, Isaac Sim → deploy on 16 embodiments.
• A C++ safety kernel — deny-by-default, screens action chunks against an Allowed Collision Matrix, deadman + E-stop.
• Observability — every run an OpenTelemetry trace, replayable in dashboard + Foxglove, foldable into a LeRobot dataset.
Everything is Apache-2.0. Feedback, issues, and contributors very welcome.
Code: github.com/OpenRAL/openral · Site: openral.com · Discord: discord.gg/3paXT2bVyB
My feedback is this you should preview your Markdown before hitting “post”. The preview is baked into the Discourse WYSIWYG editor which makes me think an LLM created this post on your behalf.
1 Like