Physical AI, like robots and humanoids, has a data problem, and the industry has not agreed on who solves it.

Robots can walk, see, and plan, yet they still fumble the basic act of handling a real object. The missing ingredient is data on how hands actually manipulate the world, and that data barely exists at scale. Teams are chasing it from every direction.

Tesla, Figure, and AgiBot collect it through teleoperation rigs and exoskeleton suits. NVIDIA generates it in simulation. Academic labs mine egocentric video for human motion. Every path trades fidelity for scale or scale for cost, and most serious players now run several at once.

Into that contest steps PSYONIC, the San Diego maker of the touch-sensing Ability Hand, with a contrarian bet that the richest manipulation data may already be flowing through a medical device worn on human arms.

At Automate 2026 in Chicago, the company is showing how manipulation data captured from a human-worn bionic hand could teach industrial robots one of the hardest skills for machines today, the simple act of handling real objects well. The work anchors a three-way collaboration with ABB Robotics and NVIDIA that PSYONIC is unveiling at the show.

For a decade, the AI race has turned on a single question.

Who has the best model? OpenAI, Anthropic , Google, Meta, and xAI have spent billions training systems on trillions of words scraped from the internet.

Larger datasets produced stronger reasoning, better code, and more capable assistants. Robotics is now exposing the limits of that playbook. A model can describe how to peel a banana, yet a robot still struggles to grip the fruit without crushing it.

PSYONIC Turns A Prosthetic A Data Engine For Robots

PSYONIC ’s core insight is that dexterity is as much a data problem as a hardware problem. Founder and CEO Dr. Aadeel Akhtar has spent years building the Ability Hand, a lightweight prosthetic with touch sensing, vibration feedback, and multi-articulating fingers that is already FDA-cleared and worn by more than 300 patients.

In chatting with Dr Aadeel, he told me,"Dexterous manipulation is as much a data challenge as it is a hardware challenge. By using the same Ability Hand with both people and robots, we can capture high-quality, real-world data on movement, touch, and grip force, then use those insights to train robotic systems more effectively."

The same device that helps amputees regain function also records something rare: high-fidelity signals on grip force, finger position, contact, and timing as a person interacts with the physical world.

That dual purpose creates a bridge between human expertise and robotic learning. Every time someone wearing the hand adjusts pressure on a slipping glass or eases a strawberry off the vine, the system captures the micro-corrections people make without thinking. Those corrections are exactly what robots lack.

PSYONIC Could Hand Robots Its ImageNet Moment

Computer vision had a turning point when ImageNet gave researchers millions of labeled images and neural networks learned to see the world at scale. Robotics has never had its equivalent. The internet holds nearly unlimited text and pictures, yet almost nothing about how human hands actually manipulate objects. Video shows the outcome of a movement and misses the tactile story underneath, the pressure, resistance, and constant adjustment that make human touch effective.

PSYONIC is working to fill that gap with NVIDIA and ABB Robotics. Earlier this year, the Ability Hand became one of the first commercially deployed dexterous hands integrated into NVIDIA Isaac Lab, an open framework for robot learning. The integration lets developers simulate, train, and validate manipulation policies using a sensorized hand that already works in the real world. PSYONIC pairs this with a real-to-real transfer approach, grounding robot training in physical interaction collected from the same hand used by both people and machines.

At Automate 2026, PSYONIC is bringing those threads together. At the show the company unveiled a three-way collaboration with NVIDIA and ABB Robotics. The three combine an open learning framework, an industrial automation platform, and a touch-enabled bionic hand to test how human dexterity data informs robotic manipulation. On the floor that means pairing the Ability Hand with ABB's GoFa collaborative robot to handle the delicate, irregular, and variable items that stump rigid systems. ABB Robotics, which SoftBank agreed to acquire for $5.3 billion in late 2025 , frames the work as core to its push toward Autonomous Versatile Robotics, machines that sense, reason, move, and handle objects in changing environments.

As Akhtar put it, "Through our collaborations with ABB and NVIDIA, we're exploring how that data can help advance robotic manipulation and contribute to the future of physical AI."

What Enterprises Betting on Robots Should Care About

The lesson reaches well past the robotics lab.

Every wave of AI has run on a scarce resource. The cloud era depended on compute, generative AI depended on internet-scale text, and physical AI may depend on human movement. The companies recording how hands touch the world now could become the infrastructure layer for the robotic economy.

Three questions belong on the table for any leader deploying robots in messy, real-world operations.

  1. Whether dexterity is the bottleneck. Traditional automation performs best when every item looks identical, and reality rarely cooperates. If your operation handles dented boxes, irregular produce, or components that shift on the line, rigid systems will keep stalling, and human-generated dexterity data is one of the few paths to robots that adapt.
  2. Who owns the data layer. The advantage is shifting from who builds the best robot to who holds the richest record of human manipulation. Prosthetics makers, wearable sensor firms, exoskeleton builders, and teleoperation startups may be sitting on proprietary maps of how people touch the world, so leaders should know whether they are buying hardware or renting access to someone else's dataset.
  3. How fast the learning curve compresses. Human-generated data lets robots observe both successful actions and the countless small corrections people make without thinking, which can cut training from years to months. That timeline decides whether automating a variable task is a someday pilot or a next-year budget line.

If language data powered the first AI revolution, dexterity data with robots may power the next, and PSYONIC is positioning itself to supply it.