The AI robotics field has made significant strides in model architecture, with innovations like diffusion models, world models, and imitation learning. But progress is increasingly gated not by the sophistication of these models but by the quality and volume of training data available to feed them.
Training a robot to perform real-world tasks requires large datasets of egocentric video, depth sensing, and, in some cases, tactile feedback, all captured in authentic environments rather than simulated ones, and the tools to collect that data at scale remain improvised, unreliable, and difficult to deploy.
The result is an industry racing to build increasingly capable machines while the infrastructure needed to teach them falls further behind.
Very few startups are building the actual recording tools that operators use in the field, but Hendrik Chiche, the cofounder and CEO of OMGrab, is among those trying to change that. “There are many startups that do data collection,” he says. “However, there are very few startups that build the tools to do data collection.” His company sits at the infrastructure layer, designing purpose-built hardware and software to make high-quality data collection scalable, reliable, and simple enough for any field operator to use.
Why Robotics Training Stalls at the Practical Level
Most industry attention and venture capital regarding robotics have focused on developing newer and more advanced models, but this means the infrastructure needed to collect training data to make sure they’re able to perform well has been treated as an afterthought. The result is a structural imbalance: teams build models with more capabilities but without high-quality data pipelines to feed them information they can work with.
The problem compounds as robotics applications are sent to real-world environments like retail, logistics, and healthcare, where different conditions will inevitably arise. A model trained on footage from a single kitchen won’t be able to adapt its settings when interacting with thousands of different ones.
For Hendrick Chiche, most robotics data collection startups today (the ones trying to solve this problem) focus on the operations layer, meaning getting human operators to gather data, instead of on the quality of the recording tools themselves. The hardware that operators actually carry into the field is often improvised. Teams commonly rely on consumer phones, off-the-shelf cameras, and SD cards to capture training footage, which means inconsistent quality, corrupted files, poor battery life, and cumbersome workflows that slow collection and frustrate the people doing the work.
The result is that data quality is wholly dependent on specific collection efforts, and they’re prone to inconsistent visual or sonic cues that affect a model’s performance. “It’s still challenging to make a really good recording system that can be worn several hours per day by an operator,” Chiche points out. “It’s a lot of small design details that make it wearable, that make the battery last longer, that make the streaming not lose information.”
OMGrab’s End-to-End Data Pipeline
OMGrab was built to fill this gap. Rather than managing operators or reselling datasets, the company designs and manufactures its own wearable recording devices paired with a cloud platform purpose-built for robotics data to create an infrastructure layer designed specifically for high-quality data collection.
The device itself has a single physical button. When powered on, it connects automatically to Wi-Fi, begins streaming, and syncs all captured video to the cloud without any additional input from the operator. A QR code in the shipping box links to setup instructions, and software updates are pushed over the air across the entire device fleet simultaneously, meaning new features reach every operator at once. The cloud platform handles storage, annotation, and fleet management, giving customers a single dashboard to monitor data collection across geographies.
By combining wearable recording devices optimized for all-day use, real-time streaming that reduces the possibility of data corruption, and a unified cloud infrastructure, OMGrab addresses the exact problems that plague improvised setups: inconsistent quality, corrupted files, and cumbersome workflows that negatively affect a model’s performance.
And underlying all of it is a commitment to field simplicity. For Chiche, if the technology can’t be deployed by a non-technical operator with minimal onboarding, across geographies and languages, it can’t scale, and that’s what the company prioritizes above all else.
“The device is simple by design,” Chiche explains. “It connects automatically to Wi-Fi. All the video gets synced to the cloud; there’s no button to press.”
Why This Requires More Than Software
One reason the data collection infrastructure layer, according to Chiche, has remained underdeveloped is that solving its demands an unusually broad set of skills. A recording device that works in a research lab is not the same as one that works reliably in the hands of a field operator in rural India. Getting from one to the other requires a team that has knowledge in mechanical engineering and CAD, embedded systems, software, DevOps, and machine learning, all disciplines that are rare to have in a startup team.
“Since it’s a hardware-and-software problem, you need skills in mechanical engineering and CAD design, skills in software engineering and DevOps, and also in machine learning,” Chiche said. “Building a tool like this requires very different skill sets.”
Chiche assembled OMGrab’s team with that gap in mind. His cofounder, Antoine Jamme, a mechanical engineer who previously built robotic arms for semiconductor manufacturing, leads hardware design. Isaac Neal, a software engineer and ML researcher Chiche met at UC Berkeley, handles the software and ML stack.
Chiche himself, who holds engineering degrees from France and UC Berkeley, understands the importance of these different sides, playing a hand in them but also leading fundraising and customer development. The company raised a pre-seed round from Founders Inc. and was already running four customer pilots by the time it officially launched in December 2025.
But while OMGrab’s team has a thorough setup, he does point out that competition is not only domestic. Chinese companies are becoming strong players in this space, benefiting from rapid access to hardware prototyping and manufacturing. Chiche acknowledges these advantages but argues that OMGrab’s integrated approach, one that seeks customer feedback to constantly iterate the entire product, allows the company to compete on design and engineering quality.
Building the Tools That Build the Robots
The robotics industry will not scale on better models alone. It needs better ways to collect and organize the data that those models depend on, and that requires purpose-built infrastructure designed from the ground up for real-world field conditions. Hendrik Chiche and OMGrab are betting that the companies that solve the data collection problem will quietly inform the trajectory of the entire field.
