Terms we use.
A short reference for the vocabulary of robotics data work as it appears across Agentuor.
Physical AI
AI systems that perceive and act in the physical world through robots or other embodied agents.
Embodied robotics
Robotic systems whose learning and behavior are shaped by their physical body and its interaction with the environment.
Episode
A single recorded attempt at a task, from start to outcome, with all sensor streams synchronized.
Multimodal annotation
Labeling that spans several data types at once — for example, camera frames, LiDAR sweeps, and joint telemetry — on a shared timeline.
Provenance
The recorded chain of evidence and decisions behind a label or recommendation: what produced it, who approved it, and under which guideline version.
Confidence score
A calibrated estimate of how likely a recommendation is correct, shown so a reviewer can prioritize attention.
Failure mode
A recurring class of model or annotation error, grouped by cause rather than by individual instance.
Sim-to-real gap
The difference in performance between a policy evaluated in simulation and the same policy on a physical robot.
Teleoperation
Human control of a robot to produce demonstration data for learning.
Expert review
Adjudication of ambiguous or high-impact annotations by a domain specialist.
Evaluation suite
A versioned set of held-out episodes used to measure model performance under defined conditions.
Managed services
A delivery model in which Agentuor staffs and executes annotation and review against agreed quality SLAs.