Embodied AI Glossary

Embodied AI: AI designed to perceive, reason and act through a physical or simulated body.
VLA: Vision-Language-Action model connecting visual and language inputs to robot actions.
Embodied Reasoning: Reasoning about physical environments, objects, goals and actions.
Robot Foundation Model: A general-purpose model intended to support multiple robot tasks or embodiments.
Sim-to-Real: Transfer of learned behavior from simulation to physical hardware.
Teleoperation: Human control of a robot used for operation, supervision or data collection.
Robot Data: Experience such as demonstrations, trajectories, sensor observations and interaction outcomes.
Manipulation: Physical interaction with objects using robot end effectors.
Dexterous Manipulation: Fine-grained object interaction requiring coordinated hands, fingers or grippers.
Whole-Body Control: Coordinating locomotion, balance and manipulation across the robot body.
Cross-Embodiment: Learning or transferring skills across robots with different bodies and sensors.
Policy: A mapping from observations and state to actions.
Imitation Learning: Learning behavior from demonstrations.
Reinforcement Learning: Learning behavior through interaction and reward signals.
World Model: A learned representation or predictive model of aspects of an environment.
End Effector: The tool or hand at the end of a robot arm.
Degrees of Freedom: Independent motion dimensions available to a robot mechanism.
Foundation Model: A broadly trained model adapted to downstream tasks.

Primary source