Teach robotsevery human skill.

The data engine for physical AI. Camera pose, hands and actions for every frame of first-person video.

#1 LaMAria mono-inertial · 750 hours open-sourced · 10,000+ hours of real-world data

workbench_0412.mp4 · illustrative
First-person view of a watchmaker's hands working on a movement with tweezers

Capture to training set.

Run the full pipeline or any single stage.
╔═ FPV LABS DATA ENGINE ═════════════════════════════════════════════════════════════════════════════╗
║                                                                                                    ║░
║   ┏━━━━━━━━━━━━━━━━━┓      ┏━━━━━━━━━━━━━━━━━┓      ┏━━━━━━━━━━━━━━━━━┓      ┏━━━━━━━━━━━━━━━━━┓   ║░
║   ┃     Capture     ┃─────▶┃     FPV-SLAM    ┃─────▶┃  Hands + Vidur  ┃─────▶┃     Evaluate    ┃──▶╫──▶ training set
║   ┃  Ego-OSCAR rig  ┃      ┃   camera pose   ┃      ┃  21 joints x 2  ┃      ┃   score every   ┃   ║░
║   ┃  or Stera, iOS  ┃      ┃   every frame   ┃      ┃  timed actions  ┃      ┃     session     ┃   ║░
║   ┗━━━━━━━━━━━━━━━━━┛      ┗━━━━━━━━━━━━━━━━━┛      ┗━━━━━━━━━━━━━━━━━┛      ┗━━━━━━━━━━━━━━━━━┛   ║░
║                                                                                                    ║░
╚════════════════════════════════════════════════════════════════════════════════════════════════════╝░
 ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

Benchmarks

LaMAria · ETH Zurich · mono-inertial

SLAM, short-walk score

FPV-SLAM
61.4
Next best
57.1
FPV-SLAM →
WGO-Bench · 100 episodes · 743 segments

Action labels, semantic F1

Vidur
0.386
Previous best
0.280
Vidur ↗
EgoExoLearn · 48 videos · 10.23 hours

Action labels, semantic F1

Vidur
0.239
Gemini 3.5 Flash
0.157
Vidur ↗
LaMAria binocular-inertial · #3 · 77.3 short-walk score
iPhone pose vs 24-camera Vicon · drift under 1 cm

Run FPV-SLAM on your footage.