Glossary
Robot training-data terms, defined.
The vocabulary buyers and models both search. Each definition is self-contained and quotable — because the difference between video and trainable data lives in these distinctions.
- Teleoperation trajectory
- A time-synchronized sequence of observations paired with the robot actions a human operator commanded — the core trainable unit for imitation learning. Video without these action labels cannot train a policy.
- 7-DoF action label
- The 6 degrees of freedom of end-effector pose (x, y, z + roll, pitch, yaw) plus 1 gripper dimension, recorded per timestep. It is the value a manipulation policy must output — the thing you are actually training.
- Proprioception
- The robot's sense of its own state — joint angles, end-effector pose, gripper width, forces — recorded alongside observations. Essential for manipulation and the primary signal for blind legged locomotion.
- RLDS / LeRobot / HDF5
- Standard container formats for robot-learning datasets. RLDS (Reinforcement Learning Datasets) and LeRobot are the community-standard episodic formats; HDF5 is the underlying hierarchical store. Delivering in these formats is what makes a dataset ingestible out of the box.
- Egocentric demonstration
- First-person human demonstration data (head-mounted or wearable capture). Useful for pretraining and affordances, but only becomes robot-trainable once an action — typically a recovered 3D hand/finger pose — is attached.
- Force / torque data
- 6-axis measurement of contact forces and moments during manipulation. Exposes shear, slip and compression that vision cannot see — essential for contact-rich assembly and publicly scarce.
- 4D Gaussian Splatting (4DGS)
- A dynamic 3D scene reconstruction technique. At Kinema it is a capture means — an intermediate we use to extract calibrated multi-view RGB-D and 3D hand/object pose — not a deliverable buyers ingest.
- Data diversity vs volume
- Past a per-scene threshold, a policy's generalization scales with the diversity of environments and objects far more than with the raw number of demonstrations (ICLR 2025, arXiv:2410.18647) — which is why bespoke, diverse capture outperforms more of the same.