Two ways to record a demonstration
A human demonstration captures a person performing a task with their own body. It provides natural perception and manipulation signal expressed in human embodiment rather than a robot's control interface. A teleoperation episode captures a task through the robot itself, so the recorded actions are already in the robot's action space.
Both are demonstrations; the difference is embodiment. That difference determines how directly the data can be used and what has to be bridged before a policy can learn from it.
Human-manipulation recordings
Datoric's multimodal human-manipulation program records synchronized egocentric and external video with RGB-D/depth, motion capture, IMU, and hand and body pose. Action segments, object and tool labels, hand-object contact, task phases, and completion states are annotated, and calibration and sensor-alignment QA keep the streams aligned.
Depth and pose make human-manipulation recordings useful for fine-grained manipulation, perception pretraining, and world-model work. The motion remains human, so a downstream robot policy still has to map from human hands to its own actuators.
Robot-teleoperation episodes
The teleoperation program records episodes performed through a robot: multi-camera and wrist-camera video paired with joint states, end-effector pose, gripper state, action commands, and trajectory logs, plus object, phase, and success/failure labels and QA reports.
Because the actions are already in the robot's action space, teleoperation data provides direct supervision for imitation learning and VLA policies. The available dataset supports capacity up to 2,000 episodes and approximately 100 to 250 hours, while the human-manipulation dataset supports capacity up to 6,000 hours.
Choosing between them
A practical rule of thumb
- Use human-manipulation data for scale, perception pretraining, depth and pose supervision, and world-model training.
- Use teleoperation data for imitation learning and VLA policies that consume actions in the robot's own action space.
- Combine them when you want broad perception coverage from human demonstrations plus embodiment-aligned fine-tuning from teleoperation.
Labels are shared, action spaces are not
Both specifications include object and tool labels, action or phase segmentation, and completion or success/failure outcomes. What differs is the action representation: human pose and contact on one side, joint states and action commands on the other. The model's expected action space is the central selection question.
Datoric