Innodata opens Vicon-equipped lab for humanoid robot training data

Innodata, a data engineering company founded in 1988, is positioning its new laboratory as a response to a shortage it sees in physical AI. The company describes large language models as having benefited from the vast amount of data on the Internet, while robots have no equivalent corpus of real-world interaction data. Its vice president of robotics and physical AI, Franklin Tanner, framed the problem as earning tokens one deliberate interaction at a time.

The lab's approach differs from inferring 3D motion out of 2D video. Innodata says it captures 3D data directly from human or mechanical bodies, and it has equipped the facility with high-precision, low-latency infrared optical tracking cameras alongside Vicon's technical consulting. Tanner acknowledged real-world data remains the gold standard but is super expensive, and pointed to simulation as a way to seed edge cases, naming NVIDIA's Cosmos environment as becoming really good at real-to-sim translation. He also flagged a gap the lab is working on: training data where humans and robots interact, and instrumenting setups so multiple agents can work together.

The stakes appear to hinge on whether buyers accept the lab's outside-in measurements as ground truth and whether its data can shrink the volume that gets discarded. One internal example cited the practice of throwing out 80% of a million hours of requested egocentric data, leaving 200,000 useful hours. For robot developers weighing their own telemetry against third-party benchmarks, that gap is likely where the lab's value will be tested.