/ GLOSSARY
The vocabulary of robot data.
Precise, working definitions for the terms that come up when you build a robot data pipeline — written for engineers who need the operational meaning, not the textbook one.
/ Teleoperation
Teleoperation
Teleoperation is the real-time remote control of a robot by a human operator, using streamed sensor feedback to close the perception-action loop.
Bilateral Teleoperation
Bilateral teleoperation reflects the forces a robot feels back to the operator's hand, closing a force loop that is far more delay-sensitive than video.
Shared Autonomy
Shared autonomy blends human input with an autonomous policy in real time, so a person and a robot control the same task jointly instead of trading it off.
Motion Retargeting
Motion retargeting maps a human operator's tracked motion onto a robot with different kinematics, and its error corrupts demonstrations if unmeasured.
End-Effector Pose
An end-effector pose is a robot tool's position and orientation in 6 degrees of freedom — frame and rotation choice both shape what a policy learns.
/ Robot learning
Imitation Learning
Imitation learning trains a robot policy to reproduce demonstrations, without a reward function — and its central failure mode is compounding error.
Vision-Language-Action Model (VLA)
A vision-language-action model maps camera images and a language instruction directly to robot actions, built on a pretrained vision-language model.
Behavior Cloning
Behavior cloning trains a robot policy with supervised learning, mapping observations to actions — and it fails in specific, well-understood ways.
Action Chunking
Action chunking predicts a short sequence of future actions per step, reducing compounding error and handling pauses in demonstrations better.
Sim-to-Real Transfer
Sim-to-real transfer means training a policy in simulation and deploying it on real hardware — and the reality gap is why that transfer isn't automatic.
/ Data pipeline
Robot Demonstration Data
Robot demonstration data pairs synchronized sensor streams with the actions a human took, structured so an imitation-learning policy can learn the mapping.
Data Flywheel
A data flywheel is the loop of deploying a policy, capturing failures, and retraining on them — it only spins for robots if collection is instrumented.
Time Synchronization
Time synchronization puts every sensor on one clock, so a dataset isn't silently pairing the wrong camera frame with the wrong joint state.
Episode
An episode is one bounded attempt at a task — a sequence of steps from a defined start to a defined end — and the unit most robot dataset formats index by.
/ Data formats
MCAP
MCAP is a self-describing, chunked, indexed log format for timestamped multimodal messages, and the default ROS 2 bag storage plugin since Iron.
LeRobot Dataset
LeRobotDataset is Hugging Face's format for robot learning: Parquet for tabular state/action, MP4 shards for video, metadata that reconstructs episodes.
RLDS
RLDS (Reinforcement Learning Datasets) stores episodes as sequences of steps on top of TFDS/TFRecord, and is the format Open X-Embodiment standardized on.
URDF
URDF is the XML format describing a robot's links, joints, and geometry for ROS. It defines kinematics but can't express closed chains or sensors natively.
Put this into practice.
Tell us what your robots need to learn. We will scope the rig, the operators, the protocol, and the first datasets — usually in one call.