Innodata opens a motion capture lab to help humanoids move more like people

While humanoid robots can be agile, the AI ​​that drives them needs more data to be able to perform useful tasks. Innodata Inc. said today it has opened a lab to generate data to train the next generation of human-like robots. The facility will also independently validate performance data generated internally by the robots.

“Physical AI is growing faster than any other segment of 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,” said Rahul Singhal, CEO of Innodata. “This framework removes that wall. We now offer physical AI companies complete end-to-end functionality, from data collection to model evaluation, that compresses development cycles and brings more capable, safer robots to market faster.”

Founded in 1988, Innodata said data and artificial intelligence “are inextricably linked.” The Ridgefield Park, New Jersey-based data engineering firm said it provides the high-quality data, evaluation frameworks and human expertise needed to build AI systems that builders and users can trust at scale.

Physical AI places new demands on data

While large language models (LLMs) have benefited from the large amount of data on the Internet, robots don’t have the same body of data about the physical world, noted Franklin Tanner, vice president of robotics and physical AI at Innodata. “Physical AI has to earn its tokens one interaction at a time, and they have to be intentional,” he said The robot report.

According to Innodata, the new research and development facility addresses the urgent need for high-quality training data for humanoids, industrial robots and other forms of physical AI. While other data providers infer 3D motion by analyzing 2D video, the company said it captures 3D data directly from the bodies themselves, regardless of whether those bodies are human or mechanical.

“There is no substitute for direct 3D motion capture,” Tanner said. “When a computer vision model tries to make sense of a 2D grid of pixels, errors inevitably creep in.”

“Our sensors are designed to register the smallest movement of each joint, which makes training more accurate and efficient,” he added. “When you train a humanoid that weighs almost 200 lbs. [90.7 kg]your reads can’t be in the ballpark. They have to be precise. And with this laboratory I am.”

Tanner cited the example of telling a robot to get a cup.

“You want to make sure he takes it the right way. So most of us take it by the handle, but you can also take it by the base,” he said. “It’s not wrong, but if it’s filled with hot liquid, you might end up burning your hand. So context matters, and it’s not always encoded even in the datasets available right now.”

Editor’s Note: Physical AI and humanoids are among the topics covered in the RoboBusiness 2026 session, which will be held on October 20 and 21 in Santa Clara, California. Register now to participate.

Data from robots in nature is the “gold standard”

Innodata said it captures real-world data and produces training data on humanoids, teleoperated hardware, wearable systems, sensor platforms, Universal Manipulator Interface (UMI) grippers and other multimodal acquisition setups. The data can then help create digital twins to generate edge cases.

“There is no substitute for real-world data,” Tanner said. “As much as possible, we collect data of an actor or a robot doing something in nature. I think that’s the gold standard. However, it’s very expensive and figuring out all the ways cups can go down a conveyor belt with the handles facing in different directions is not efficient.”

“The crux is how do we create enough real-world data and then run a simulation with that?” he asked. “Simulators are getting pretty good at taking data from the real world, doing the conversion from real to simulated, and then modifying the environment to handle some of these little combinatorial explosions of different parameters. NVIDIA’s Cosmos environment is actually getting really good at this.”

There are also limitations to real-world data. Tanner said human responses to teleoperation or egocentric data acquisition can be incomplete. If a person in a data collection scenario drops a knife, for example, they might step back and turn off the device thinking they’ve made a mistake, he said.

“We don’t currently have training data where humans and robots interact,” Tanner noted. “It’s an area we’re working on in the lab: equipping it so multiple agents can work together. We need larger, more customized datasets.”

Vicon captures submillimeter motion for Innodata

Innodata developed the lab with Vicon, a leader in the motion capture industry, who continues to provide technical consultancy. The New Jersey lab is equipped with high-precision, low-latency infrared optical tracking cameras that measure motion down to the submillimeter level—a degree of precision that methods like wearable IMU (inertial measurement unit) sensors and single-camera (monocular) video analytics typically can’t match.

“A lot of motion capture companies have discovered humanoid robotics lately,” acknowledged Andrew Knox, CEO of Vicon. “But teams building the most capable robots keep coming to the same conclusion: If the training data is rough, the robot will be rough, too.”

“Vicon sets the standard against which motion data is measured,” he said. “Our systems capture people, robots and the objects they manipulate in the same space, with submillimeter precision and millisecond latency.”

“This provides an independent ground truth, not an estimate,” Knox explained. “That’s why leading humanoid developers, frontier AI labs, and research universities continue to choose Vicon, and why Innodata’s lab is well positioned to become a benchmark for the entire physical AI industry.”

Innodata customers can purchase data and send their robots

Innodata customers will be able to purchase motion capture data in standard packages or commission custom projects. The company said it can produce training data for a wide range of physical AI platforms and redirect motion data from one platform to another as needed.

Customers can also submit robots for evaluation to confirm that they perform according to customer specifications. Innodata experts can measure a robot’s movements and responses in a variety of programmed scenarios, including those in which robots and people interact.

Innodata also provides safety assurance services to validate robot performance. The company said its physical AI experts apply the same rigorous, multi-step quality processes that it has honed over decades of providing premium data for mission-critical use cases in finance, government and healthcare, and, more recently, for frontier AI labs.

Because the lab’s finely calibrated cameras observe each robot from the outside, they provide external (“exocentric”) control over the robot’s internal (“egocentric”) telemetry, which is often noisy. The resulting measurements can be used to validate a developer’s internal data and performance claims or to provide independent, third-party benchmarks for reliability and security, Innodata said.

“We need to figure out how to do more with less data,” Tanner acknowledged. “In my opinion, this is going to be one of the hallmarks of the next evolution of robots. We have customers who say, ‘Just give me a million hours of self-centered data doing everything under the sun.’”

“Well, yes, but by my metrics you will eliminate 80% of that data,” he said. “So why do you want a million hours? Why don’t you just want 200,000 hours that will actually be useful for your VLA? This is one of the areas our team is looking at with a university partner.”