Data model

The data structures your client receives. Units are meters; rotations are quaternions (x, y, z, w); positions are world space unless noted.

Skeleton

A tracked person for one frame.

Field	Type	Notes
Id	string	Per-frame/session instance id
EntityId	string	Stable identity (set during entity registration)
AnchorId	string	Anchor the skeleton is positioned against
Positions	float[25 × 3]	Joint positions (x, y, z) in meters, world space
Confidence	float[25]	Per-joint confidence 0–1
CaptureTime	double	Unix timestamp (s, fractional) — use to sync streams
HeadLocalRotation	quaternion	Device-local head orientation
LeftHand* / RightHand*	bytes	Hand joint positions/rotations (when hands enabled)
Controller info	—	VR controller state, when present

Joints (25)

Nose · Neck · Shoulders (L/R) · Elbows (L/R) · Wrists (L/R) · Hips (L/R) · Knees (L/R) · Ankles (L/R) · Eyes (L/R) · Ears (L/R) · Toes / heels (L/R).

Hands add 21 key points each. A full body-plus-hands point set is 25 + 21×2 = 67 points — this is the positions_67 layout you'll see in exports.

Tracked objects

Type	Key fields
Marker	Id, Type, Position, Radius
Object	Id, Name, Position, Rotation, MarkerIds[]

Objects are composed from markers (e.g. a ball built from a colored marker).

Camera

Field	Notes
Id	Camera identifier
Intrinsic	3×3 camera matrix
Extrinsic	4×4 pose
Distortion	Lens coefficients
Width / Height / FrameRate	Stream properties

Spatial

Type	Fields
Anchor	AnchorId, Position, Rotation, ParentId, Visible
Guardian	GuardianId, Points[], Height (play-space boundary)

tip

EntityId is the field you want for "is this the same person as before?" — Id can change frame to frame; EntityId is stable once the person is registered.