Concepts

Glossary

Term Meaning
ORA Open Robot-Use Agents; the Python import and CLI are ora, the distribution is ora-agent.
native_api policy The policy kind (policy.kind: native_api) in which ORA itself calls the model API (OpenAI-compatible or Anthropic) and supplies the prompts, tools and context. The other kind, scripted, runs fixed actions for infrastructure checks. "Native" in names such as native_full_explicit_motion_direct_v4 refers to this policy kind, not to an agent core: native_api is only the name of the policy kind. The run's native/ evidence directory, the native_slots_* compositions and the ora.native-* schemas keep the word for the same reason.
Agent core The implementation that runs a native_api policy's agent loop: provider calls, request assembly, context management and agent-side evidence. ORA has one, Pi, and every plan records it in policy.agent_core (agent core).
Pi ORA's agent core: TypeScript in ora-pi/ on the pinned Pi agent packages, run by Node.js. Its terminal UI is the one ora interactive, ora attach and ora play open; in free play the agent is Pi itself, working as a coding agent.
Native core The Python agent core ORA started with (ora.agent.native). Removed from ORA; it remains only in the robot-harness repository. A plan or setting that names it (policy.agent_core: native, ORA_AGENT_CORE=native, --agent-core native) is refused; its sealed runs still evaluate, open in the viewer and replay.
RunSpec Portable task, provider, runtime and budget inputs.
RunPlan The resolved, content-addressed execution contract, including source and asset identities.
Baseline A maintained composition selected by the default resolver.
ToolSurface The advertised tools, schemas, capabilities and state-effect contract.
Prompt bundle Ordered project-owned policy layers used with an embodiment and environment.
Embodiment Public robot geometry, frames, units and control semantics.
Supervisor Trusted authority that checks and dispatches robot mutations.
Receipt Measured execution feedback returned by the environment/controller.
Qualification Evidence that a specific infrastructure dependency satisfies a stated contract.
Official verdict The benchmark's success boundary; never the model's self-assessment.
Seal Hash-bound immutable inventory of evaluator-valid run evidence.
Context artifact Recoverable episode-local tool/media evidence, not cross-episode Task Memory.
Free play ora play: a qualified scene with no task. The operator gives instructions one at a time; the session is sealed but never scored (free play).
Practice The plan partition partition: practice (the default is evaluation): development episodes that may not use any benchmark's evaluation episodes. Practice runs are never evaluation results (practice and evaluation partitions).
Operator view A camera image for the human watching a run, such as a play scene's third-person camera. It is recorded as evidence and shown in ora web, never given to the policy (operator view).
Replay Watching a finished run again from its sealed evidence: ora watch in the terminal, or ora replay on ora web's page or as an exported video (replay a finished run). A replay reads the evidence and never runs the model or the simulator.
Experiment family, generation Optional lineage labels for a study: a family such as EXP90A names the question, and a generation FAMILY/branch/gNNN names one treatment (experiment identity). Example cohort names such as libero-example-g001 borrow the gNNN suffix.

Source: docs/glossary.md