A data company just built a lab whose only job is teaching robots to move like us

Innodata, a data engineering company founded in 1988, opened a motion-capture lab in Ridgefield Park, New Jersey, built in partnership with Vicon, a company known for high-precision motion-capture camera systems. The lab captures 3D motion data directly from human and robot bodies, rather than trying to reconstruct movement from flat video footage.

• Accuracy: sub-millimeter precision, with millisecond latency.
• Capture methods: teleoperated hardware, wearable systems, and specialized grippers.
• Robot capacity: has handled robots weighing nearly 200 pounds.
• Services: training data generation, independent performance validation, safety assurance, and cross-platform motion retargeting.

The catch: video looks like an easy, cheap way to gather training data, since cameras are everywhere already. But Franklin Tanner, Innodata's VP of Robotics and Physical AI, explained why that approach falls short: "When a computer vision model tries to make sense of a 2D grid of pixels, mistakes inevitably creep in."

A flat video frame doesn't actually contain real depth information, a model has to guess at it. Direct 3D motion capture, by contrast, records precise spatial position and movement in real time, without that layer of inference and its accompanying errors.

In real life it's the difference between describing how someone walked based on a single photo versus actually watching them walk with a tape measure running the whole time.

Why it matters: Innodata CEO Rahul Singhal put the core issue plainly: "Physical AI is growing faster than any other segment in AI, but every robotics team hits the same wall: there isn't enough real-world interaction data, and what exists is expensive and slow to produce."

That's a fundamentally different problem than the one facing large language models. Background: an LLM can train on enormous amounts of text that already exists across the internet. A humanoid robot learning to grip a cup or walk across an uneven floor has no equivalent pre-existing dataset; that data has to be physically generated, one recorded motion at a time, which is exactly the bottleneck this lab is built to chip away at.

Vicon's managing director, Andrew Knox, summed up why the accuracy bar matters so much: "If the training data is approximate, the robot will be too." A robot that learns from slightly-off training data doesn't fail gracefully, it moves slightly wrong in ways that compound as tasks get more complex.

Who's affected: that's part of why the lab also offers independent robot performance validation as a separate service. A company can send a finished robot to have its real-world movement measured externally, rather than relying only on its own internal testing, giving outside verification that a robot actually performs the way its maker claims.

That kind of third-party check matters more as humanoid robots move from investor demo days toward actual paid deployments. A robotics company claiming its machine can safely work alongside people carries a lot more weight when an outside lab, not the company itself, is the one measuring how it actually moves.

Innodata hasn't publicly named which robotics companies are using the facility. The lab is positioned broadly for physical AI companies building humanoid robots, industrial robots, and other robots meant to operate in real-world physical environments, rather than serving one exclusive customer.

Background: that positioning fits Innodata's existing business model. The company has spent decades providing data and AI training services across industries, and this lab extends that same specialized data-services approach into robotics specifically, rather than Innodata building or selling robots of its own.

Background: Innodata isn't a robotics startup chasing a trend, it's a data engineering company that's been in business since 1988, historically providing data labeling, annotation, and AI training services across industries. The company has grown sharply as AI demand surged more broadly: it reported $251.7 million in 2025 revenue, up 48% year over year, and management has guided for at least 40% growth in 2026.

That growth has come with real customer concentration risk. In 2025, a single customer accounted for roughly 58% of Innodata's total revenue, and by the second quarter of 2026, its two largest customers combined still made up about 71% of quarterly revenue. Diversifying into new specialized services, like a dedicated motion-capture facility for physical AI, is a plausible way to reduce that dependence on just one or two massive clients over time.

Why it matters: framed that way, the motion-capture lab isn't just a robotics curiosity. It's a bet by an established AI-data business that physical AI and humanoid robotics will need the same kind of specialized data infrastructure that generative AI models already rely on, and that being an early, credible provider of that infrastructure is worth the investment now.

By the numbers: physical AI, the broader category covering humanoid robots, industrial robots, and other embodied AI systems, has been described by Innodata's own CEO as growing faster than any other segment of AI. That claim lines up with a wave of humanoid robot companies raising large funding rounds and racing to get working robots into warehouses, factories, and eventually homes over the next few years.

Each of those robots needs training data specific to physical movement, grasping, balance, and human interaction, data that can't simply be scraped from the internet the way text or images can. A specialized lab that multiple robotics companies can use, rather than each one building its own in-house motion-capture facility from scratch, is a practical way to meet that demand without every robotics startup duplicating the same expensive infrastructure.

• You won't interact with this lab directly, but it's the kind of infrastructure that shapes how capable and natural-looking the humanoid robots you do eventually see in stores, warehouses, or homes actually become.
• It's a sign the humanoid robotics industry is maturing past flashy demos. Specialized data and validation infrastructure like this tends to show up once an industry is serious about real deployment, not just prototypes.
• Independent robot performance validation matters if you're evaluating robotics companies' claims. A third party actually measuring whether a robot performs as advertised is a meaningfully different kind of evidence than a company's own marketing demo.
• Expect more infrastructure like this to emerge, as robotics companies increasingly need specialized data and testing services rather than building everything in-house themselves.