Demonstration Data

The recorded observations and actions from a human performing a task, used as raw material for robot imitation learning.

Demonstration data is the raw material of imitation learning: a human performs a task once, through teleoperation, kinesthetic teaching (physically guiding the robot's arm), or a handheld capture device, and the system records, at a fixed rate, each frame's observation — camera images, joint angles, and other proprioceptive state — together with the matching action. One complete run from start to finish is stored as a trajectory, usually called an episode in a dataset; an episode recorded while the robot executes on its own does not count as a demonstration. A policy network learns exactly this mapping: given this observation, do this action. Quality drives results directly — whether the motion is consistent, the camera stays fixed, and the object stays in view all matter. A single task can reach a usable policy from a few dozen demonstrations, while general-purpose VLA (vision-language-action) models need pretraining on thousands of hours of demonstrations across many tasks and scenes.

ExampleIn the ALOHA paper, most bimanual tasks used just 50 demonstrations (roughly 10 to 20 minutes of data per task); trained with the ACT algorithm, the task of inserting a battery into a remote control reached a 96% success rate. The LeRobot tutorials recommend recording at least 50 demonstrations for a beginner grasping task.

Also called

Demonstrations, Expert Demonstration, Demo Data

Related

Imitation Learning · Behavior Cloning · Teleoperation · Observation-Action Pair · Episode · Real-Robot Data

Sources

ALOHA: Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
Zhao et al. 2023: Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (arXiv 2304.13705)
LeRobot Docs: Imitation Learning on Real-World Robots