Humanoid robots learn from people first — we collect the human demonstration data that trains VLA models.

GELLO / leader-arm teleoperation across pick-and-place, folding, pouring and assembly tasks.

VR teleoperation: whole-body manipulation and locomotion — the most valuable embodiment-specific type for humanoid companies.

Handheld gripper with camera in real environments — cheapest type to collect, scales beyond the lab: apartments, cafés, stores.

First-person capture of household and workplace tasks — for pre-training VLA and video models, sold by volume.

One task captured synchronously from 4–6 cameras plus an egocentric view — for representation learning, 3D reconstruction and cross-view learning.

Failed attempts plus operator error correction — rare and in demand: most open datasets contain only successful episodes.