Tao Chen

Most robots can lift a car door. Very few can pick up a coin off a table. Tao Chen has spent his career on the second problem, the harder one, the one that decides whether robots ever leave the factory cage and walk into your kitchen.

Today he is the co-founder and CEO of Dexmate, a robotics company in Santa Clara building general-purpose machines with hands that actually work. The flagship is Vega, a dual-armed mobile robot on an omni-directional base. It stands about 171 centimeters tall, folds its torso upward to reach 2.2 meters for overhead jobs, lifts 15 kilograms per arm, and runs for more than ten hours on a charge. His team reportedly took it from concept to a shippable product in under six months. That pace is the whole thesis: the AI is finally good enough, so the bottleneck moved to building.

The conviction underneath all of it is narrow and stubborn. Robots will only matter when their fingers can do the small, fiddly, deeply human things, threading a spray bottle, flipping food in a wok, opening a door whose handle nobody warned the robot about. Chen calls the goal plug-and-play: a robot you set down in a room that was built for people, that gets to work without anyone rebuilding the room for it.

From a mechanical engineering desk to a best-paper podium

Chen entered Shanghai Jiao Tong University to study mechanical engineering and automation. Somewhere in his senior year the field tilted under him and he pivoted into AI, the same pivot that would later define his co-founder. He graduated in 2016 with a thesis ranked in the top one percent of his class, then spent time as a research engineer at a Shanghai robotics startup working on SLAM, object detection, and reinforcement learning, the gritty plumbing of getting a machine to understand where it is and what it sees.

Then came the American leg. A master's at Carnegie Mellon's Robotics Institute under Abhinav Gupta. A Ph.D. in electrical engineering and computer science at MIT's CSAIL under Pulkit Agrawal. His early doctoral work taught quadrupeds to climb over obstacles. Around 2020 he made a deliberate turn toward dexterous manipulation, betting that hands, not legs, held the bigger commercial prize. The bet paid an early dividend in 2021, when his paper on general in-hand object reorientation won the Best Paper Award at CoRL, one of the field's most selective conferences.

The trick that made people pay attention

The headline result, later expanded into a Science Robotics paper called Visual Dexterity, was a robot hand that could reorient objects in its grasp, rotating and repositioning them mid-air. In some setups it managed this largely by feel, without leaning on cameras at all. To a non-roboticist that sounds modest. To anyone who has tried to program a machine to turn a key, it is close to magic. Touch is high-dimensional, noisy, and unforgiving, and most of the field had quietly agreed to avoid it. Chen leaned in.

He also helped build PyRobot, an open-source robotics framework that collected thousands of stars on GitHub, and Dexmate later pioneered controlling a dexterous hand through an Apple Vision Pro, teleoperating fingers with your own, then open-sourced that too. There is a pattern here: solve the hard sensing problem, then give the tools away.

The flywheel, and the 90 percent that is noise

Dexmate's strategy hangs on what Chen calls a data flywheel. Rather than collecting training data in a straight, expensive line, the company uses the data it already has to generate more, faster, blending simulation, learning from human video, and real-world demonstrations instead of picking a single religion the way some rivals do. Chen is clear-eyed about the cost: reinforcement learning can spit out billions of data points an hour, and by his estimate less than ten percent of it is actually meaningful. The craft is in knowing which ten percent.

On hardware, his design taste is specific. The best dexterous hand, he argues, is a rigid skeletal core wrapped in soft material, hard enough to grip with force, soft enough to feel. Dexmate co-designs the mechanics, sensors, control policies, and AI models all at once rather than treating hardware and software as a relay race. It is the kind of decision that sounds obvious in a sentence and is brutally hard in practice.

The co-founder who runs marathons

Chen started Dexmate in 2024 with Yuzhe Qin, whose resume reads like a mirror image: same university in Shanghai, same pivot from mechanics to AI, frontier manipulation research at UC San Diego. One of the quiet surprises of the early days, Chen has said, was how readily hardware veterans with ten and twenty years of experience joined a founding team whose average age was under thirty. That talent, more than any single algorithm, is what let Vega go from sketch to product so fast.

The company is an NVIDIA Inception startup and has reportedly raised around $33 million from investors including LG Technology Ventures and Epsilon Ventures. Vega is available to preorder at roughly $89,999 with a $999 deposit, aimed at manufacturing, logistics, retail, and eventually the home. The aspiration Chen states plainly: versatile, nimble, reliable robots that take the tedious, dangerous, and exhausting work off human hands. The marathon, in other words, is still being run.