Innodata has opened a new lab to capture motion for accurate data to train AI for robotics. Source: Innodata
While humanoid robots can be agile, the artificial intelligence driving them needs more data to be able to conduct useful tasks. Innodata Inc. today said it has opened a laboratory to generate the data for training the next generation of human-like robots. The facility will also independently validate the performance data that robots generate internally.
“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,” stated Rahul Singhal, CEO of Innodata. “This facility removes that wall. We now offer physical AI companies a complete end-to-end capability — from data collection through model evaluation — that compresses development cycles and gets more capable, safer robots into the world sooner.”
Founded in 1988, Innodata asserted that data and AI “are inextricably linked.” The Ridgefield Park, N.J.-based data engineering company said it provides the high-quality data, evaluation frameworks, and human expertise required to build AI systems that can builders and adopters can trust at scale.
Physical AI poses new data demands
While large language models (LLMs) benefitted from the vast amount of data on the Internet, robots do not have the same corpus of data on 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 deliberate,” he told The Robot Report.
The new research and development facility addresses an urgent need for high-quality training data for humanoids, industrial robots, and other forms of physical AI, according to Innodata. While other data providers infer 3D motion by analyzing 2D video, the company said it captures 3D data directly from bodies themselves — whether those bodies are human or mechanical.
“There’s just no substitute for direct 3D motion capture,” Tanner stated. “When a computer vision model tries to make sense of a 2D grid of pixels, mistakes inevitably creep in.”
“Our sensors are designed to register the tiniest motion of every joint, which makes training more accurate and efficient,” he added. “When you’re training a humanoid that weighs almost 200 lb. [90.7 kg], your readings can’t be in the ballpark. They need to be precise. And with this lab, they are.”
Tanner cited the example of telling a robot to pick up a mug.
“You want it to make sure that it’s picking it up in the right way. So most of us pick it up by the handle, but you can also pick it up from the base,” he said. “That’s not wrong, but if it is filled with hot liquid, you may end up burning your hand. So the context matters, and that’s not always encoded in even the data sets that are out there right now.”
Editor’s note: Physical AI and humanoids are among the session topic tracks at RoboBusiness 2026, which will be on Oct. 20 and 21 in Santa Clara, Calif. Register now to attend.
Register now and help us celebrate 20 years of RoboBusiness!
Data from robots in the wild is the ‘gold standard’
Innodata said it captures real-world data and produces training data across humanoids, teleoperated hardware, wearable systems, sensor rigs, Universal Manipulator Interface (UMI) grippers, and other multimodal capture setups. The data can then help build digital twins to generate edge cases.
“There’s no substitute for real-world data,” said Tanner. “As much as possible, we collect data of an actor or a robot doing something in the wild. I think that’s the gold standard. However, that’s super expensive, and figuring out all the ways mugs can come down a conveyor with their handles facing different directions is not efficient.”
“The crux of it is how can we create enough real-world data and then seed a simulation with that?” he asked. “Simulators are getting pretty good with taking real-world data, doing the real-to-sim translation, and then permuting the environment to take care of some of these little combinatoric explosions of different parameters. NVIDIA‘s Cosmos environment is actually becoming really good at that.”
There are also limitations to real-world data. Tanner said that human responses for teleoperation or egocentric data capture can be incomplete. If a person in a data-collection scenario drops a knife, for instance, he or she might take a step back and turn off the device thinking they made a mistake, he said.
“We don’t have training data right now where humans and robots are interacting,” noted Tanner. “It’s one area we’re working on in the lab — instrumenting it so multiple agents can work together. We need larger and custom data sets.”
Vicon captures sub-millimeter motion for Innodata
Vicon says its cameras can capture data with submillimeter accuracy and millisecond latency. Source: Innodata
Innodata developed the lab with Vicon, a motion-capture leader that continues to provide technical consulting. The New Jersey lab is equipped with high-precision, low-latency infrared optical tracking cameras that measure movement down to the sub-millimeter level — a degree of accuracy that methods such as wearable inertial measurement unit (IMU) sensors and single-camera (monocular) video analysis typically cannot match.
“Plenty of motion-capture companies have discovered humanoid robotics lately,” acknowledged Andrew Knox, managing director of Vicon. “But the teams building the most capable robots keep reaching the same conclusion: If the training data is approximate, the robot will be too.”
“Vicon sets the standard that motion data is measured against,” he said. “Our systems capture people, robots and the objects they handle in the same space, with sub-millimeter accuracy and millisecond latency.”
“That gives you independent ground truth, not an estimate,” explained Knox. “It’s why the leading humanoid developers, frontier AI labs and research universities keep choosing Vicon, and why Innodata’s lab is well placed to become a reference point for the whole physical AI industry.”
Innodata customers can buy data, send in their robots
Innodata customers will be able to purchase motion-capture data in off-the-shelf packages or to commission custom projects. The company claimed that it can produce training data for a wide range of physical AI platforms and retarget motion data from one platform to another as needed.
Customers can also send in robots for evaluation to confirm they perform to customer specifications. Innodata’s experts can measure a robot’s movements and responses in a variety of scripted 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-stage quality processes that it has honed over decades of delivering 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 an external (“exocentric”) check on 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 trust and safety, said Innodata.
“We need to figure out how to do more with less data,” acknowledged Tanner. “In my opinion, that’s going to be one of the hallmarks of the next evolution with robots. We have customers that say, ‘Just give me a million hours of egocentric data doing everything under the sun.'”
“Well, yeah, but you’re going to throw out, according to my metrics, 80% of that data,” he said. “So why do you want a million hours? Why don’t you just want the 200,000 hours that’s actually going to be useful for your VLA? This is one of the areas our our team is looking at with a university partner.”