Innodata opened a motion-capture lab in Ridgefield Park, New Jersey, built to produce training data for humanoids and other physical AI systems. The facility uses Vicon cameras that track motion in infrared at sub-millimetre accuracy. The company is targeting the industry's bottleneck: a shortage of data from real interactions.
Key takeaways
- Location: Ridgefield Park, New Jersey
- Vicon infrared optical tracking, sub-millimetre accuracy, millisecond latency
- 3D motion measured directly off human and robot bodies, not inferred from 2D video
- Offering: bespoke data for a specific platform plus off-the-shelf packages
- Also independent robot performance validation, safety assurance and real-to-sim translation
The problem the lab is meant to solve
Language models got the internet. Robots have no equivalent corpus, because every interaction with the physical world has to be performed separately.
It is the most commonly named barrier in physical AI.
Physical AI has to earn its tokens one interaction at a time, and they have to be deliberate.
Franklin Tanner, Vice President of Robotics and Physical AI at Innodata.
The comparison to tokens lands. In text a token costs a fraction of a cent. In robotics the unit is a motion performed by a real body.
Why not video
The cheapest route to motion data is analysing 2D footage and computing 3D positions from it. Innodata argues that method introduces error at the input stage, because a joint's position is inferred rather than measured.
The alternative is optical motion capture?Motion capture: Measuring body movement in three dimensions using cameras and markers. It returns each joint's position without inferring it from an image., where infrared cameras track markers placed on a body or on the robot.
| Criterion | Positions inferred from 2D video | Optical motion capture |
|---|---|---|
| Source of 3D position | inference from the image | direct measurement |
| Error | accumulates at the inference step | below one millimetre |
| Latency | depends on processing | milliseconds |
| Entry cost | low | high, needs a hall and hardware |
Per the company, the sensors register the tiniest motion of every joint, which makes training more accurate and more efficient.
More than data collection
The offering goes beyond recording. Alongside data tailored to a specific platform and off-the-shelf packages, the lab sells independent robot performance validation and safety assurance.
There is also real-to-sim translation, moving captured motion into simulation. That is the reverse direction from classic sim-to-real.
Why it matters
If motion-data quality becomes something you buy externally, the advantage shifts from owning a data-collecting fleet to knowing what to do with the data.
That reorders a sector where the count of deployed machines has been the main currency.
What's next
- Innodata has not published pricing or lab capacity measured in recording hours
- The company claims a full path from data collection to model evaluation as a single service
- Independent robot performance validation remains an area with no accepted industry standard
Sources
- The Robot Report — Innodata opens motion-capture lab to help humanoids move more like people
- Innodata — Physical AI
