The Robotics Data Glossary.

A 6DoF pose describes the position and orientation of a rigid body in three-dimensional space. The six degrees are translation along three axes and rotation around three axes. In robotics, the pose may describe an end effector, camera, object, robot base, human head, tool, or sensor frame.

Behavior cloning is a form of imitation learning that trains a policy to predict an expert's action from an observation or state using supervised learning. In robotics, the training set usually contains synchronized observation-action pairs or sequences collected through teleoperation, kinesthetic teaching, scripted control, or another expert policy.

Dataset provenance is the traceable record of where data came from, how it was collected, which permissions apply, what transformations and quality decisions occurred, and which release contains the result. It lets a buyer connect a delivered asset back to its source, processing history, approval state, and intended use.

Sensor data captured from the viewpoint of the person, robot, or agent performing the task.

Capture, timing, calibration, task, annotation, quality, rights, and release context attached to egocentric assets.

Data captured by a camera or sensor observing the actor and environment from an external viewpoint.

Human demonstration data is a recorded example of a person performing a task, packaged so a learning system can study the observations, motion, task sequence, object interactions, and outcome. It may include egocentric video, external views, pose, hand state, inertial signals, audio, tactile data, annotations, and rights metadata.

Imitation learning is a family of methods that learn a policy from expert demonstrations. In robotics, demonstrations usually contain state-action or observation-action sequences produced by a human operator, another controller, a robot expert, or a transformed human example. The learner uses these examples instead of discovering the task only through trial and reward.

Multimodal robotics data combines two or more distinct signal types from the same task or episode, such as RGB, depth, audio, IMU, force, tactile, robot state, actions, and language. The modalities become useful together only when their timestamps, coordinate frames, calibration, validity, and episode context are preserved.

Temporal alignment of image frames with inertial measurements, including offsets, drift, calibration, and validity.

Robot action data records what command a robot policy, controller, or human operator issued at each step of an episode. Depending on the system, actions may be joint targets, joint velocities, torques, end-effector pose changes, gripper commands, base motion, or higher-level skill instructions.

Robot proprioception is the sensing of the robot's own configuration and motion. It commonly includes joint position, joint velocity, effort or torque, gripper state, base motion, inertial measurements, and controller status. These signals tell a policy where the robot is and how its body is changing.

Robot teleoperation is the control of a robot by a human operator through an interface such as a joystick, leader arm, VR controller, motion tracker, keyboard, or wearable device. In robot learning, teleoperation is also a collection method for recording synchronized observations, robot state, operator commands, actions, and task outcomes.

A robot trajectory is a time-indexed sequence describing how a robot or one of its components moves. It can contain joint positions, velocities, accelerations, efforts, end-effector poses, gripper state, base motion, or control targets. Timing is what distinguishes a trajectory from an unordered set of poses or a geometric path.

Observations, state, actions, outcomes, and metadata packaged to train or evaluate robotics models.

Sensor fusion is the process of combining measurements from multiple sensors to estimate a state or property more reliably than one sensor can alone. Robotics examples include fusing camera and IMU data for motion, wheel odometry and GNSS for localization, or vision and force data for contact-aware manipulation.

Time-aligned first-person and external observations of the same task or event.

Data that passes the defined synchronization, structure, quality, rights, packaging, and loader acceptance gates.

A vision-language-action model is a robot policy that conditions on visual observations and a language instruction, then produces actions that a robot can execute. Unlike a vision-language model that answers with text, a VLA must connect what the robot sees and is asked to do with a defined control space.