OpenRAL: Open-source agentic harness unifies physical AI with ROS 2-native typed contracts

OpenRAL, the Robot Agentic Layer, is a new open-source runtime designed to solve the fragmentation in embodied AI. Every robot SDK, VLA, and sensor speaks its own dialect, forcing developers into one-off glue code. OpenRAL provides a single typed contract via Pydantic v2 schemas on top of ROS 2, tf2, MoveIt 2, Nav2, and ros2_control. It introduces rSkills—robot skills packaged like models on the Hugging Face Hub with typed manifests, weights, and reproducible evaluation. These skills encompass VLA policies (30–200 Hz), perception detectors, scene VLMs, reward models, and ROS actions, all hot-swappable.

Safety is a first-class concern: a C++ safety kernel enforces deny-by-default action screening against an Allowed Collision Matrix, with deadman and E-stop support. The system separates fast, reactive control from slow LLM-based reasoning—the reasoner dispatches skills via typed tool-calls while low-level policies run at high frequency. Perception pipelines lift 2D detections into a live, tf2-aware world state with spatial memory. Every run is traced with OpenTelemetry, replayable in dashboards or Foxglove, and foldable into LeRobot datasets. OpenRAL supports simulation-to-real transfer across sixteen robot embodiments and is fully open-source under Apache-2.0.