THE SPEC SHEET

01 CAPTUREOff-the-shelf rig: RGB, LiDAR depth, hand pose.
CameraiPhone Pro, 60fps RGB
DepthLiDAR, metric scale
PoseARKit 6DoF hand and head
MountHead or chest
Extra viewsFixed cameras, partner set
TactileMulti-zone force, palm and fingertips. Shear. Wireless, time-synced.ROADMAP
HardwareOff the shelf. Replicable.

Validated end to end on HOT3D, EgoDex, OpenTouch. 242 tests passing. Transcode fidelity 0.00 reconstruction error, measured on public ground-truth data.

WHAT IT TRAINS

Built for the model classes that learn from human demonstration.

ROBOT FOUNDATION MODELS

Human episodes co-train beside robot data; hand pose stands in as the action stream.

DEXTEROUS HAND POLICIES

Hand pose retargets to multi-finger hands. The closest match to what we record.

WORLD MODELS

Egocentric footage is the mid-training stage between web video and robot data.

IMITATION POLICIES, VIA RETARGETING

Human motion converts to robot trajectories for teleop-native trainers.

Teleop-native policies need robot action labels — human data reaches them through retargeting, not directly.