Glossary

Terms as used in OpenRAL.

• VLA — Vision-Language-Action model. Maps (images, language, state) → action[chunk]. Examples: π0, GR00T N1.x, SmolVLA, OpenVLA-OFT.
• WAM — World Action Model. Generative simulator used for mental rollouts and failure anticipation. Examples: Cosmos Predict, Genie 3, IRASim, UnifoLM-WMA-0.
• rSkill — A packaged, capability-tagged unit of robot behavior (sigstore signing planned, not yet implemented). One HF Hub repo. Loaded by rSkill.from_pretrained(...).
• rSkill catalog — The OpenRAL org on the Hugging Face Hub, tagged rskill; openral rskill search is how you find one — rskills/ in this repo is only the worked-examples/fixtures subset, not the catalog.
• HAL — Hardware Abstraction Layer. The openral_hal.HAL Protocol; Python adapters live in python/hal/; every robot runs the one manifest-driven ROS lifecycle node in packages/openral_hal_node/, under the ROS node name openral_hal_<robot_id> (a node name, not a package).
• WorldState — A typed snapshot consumed by Skills and the Reasoner. Backed by tf2 + sensor topics.
• Reasoner / S2 — The slow planning loop (LLM → typed tool calls).
• Skill / S1 — The fast policy loop (VLA or scripted, action-chunked).
• Cerebellum / S0 — The C++ realtime layer below S1, where applicable (humanoids).
• Embodiment tag — A short string (e.g., g1, so100_follower, aloha, widowx) that maps a skill to a robot's heads / action layouts.
• Trace — An OTel span tree + LeRobotDataset row capturing one execution end-to-end.
• Replanning ladder — The bounded sequence: retry → param-tweak → substitute → goal-replan → human-handoff.