There is a genre of robotics video where a humanoid machine does a backflip, and the internet decides the future has arrived. Dexmate, a robotics company in Santa Clara, appears to have watched all of those videos and concluded, reasonably, that a backflip is not a business. A robot that can reliably grab a box off a shelf - that is a business. So Dexmate built Vega, and gave it wheels.
This is a slightly heretical choice. The prevailing bet in humanoid robotics is that robots should have legs, because humans have legs and the world was built for humans. The counterargument, which Dexmate is quietly making, is that legs are enormously hard, frequently fall over, and mostly serve to impress people who are not going to buy the robot. An omnidirectional wheeled base is boring. It also does not tip over while holding fifteen kilograms of someone's inventory.
Founded in 2024, Dexmate was started by a trio with the kind of resumes that make the choice to skip legs look deliberate rather than lazy. Tao Chen, the CEO, holds a PhD from MIT, has published roughly twenty-two papers in AI and robotics, and picked up a best-paper award at CoRL - which is to say he knows exactly how hard bipedal locomotion is, and decided it was not the problem worth solving first. His co-founders, Yuzhe Qin (a UCSD PhD and simulation specialist) and Chongyang Wang (MIT, with a decade of operational experience), round out a team whose collective pedigree reads like a robotics faculty lounge: MIT, UCSD, CMU, NVIDIA, Toyota Research, and more.
The thing they built, Vega, is a general-purpose mobile robot. It stands about 171 centimeters tall, weighs roughly 135 kilograms, and moves at four kilometers an hour. Its defining trick is a foldable torso: it can compress down small enough to roll through a standard doorway, then unfold and stretch its arms to reach 2.2 meters - seven feet and change - to handle the overhead tasks that give warehouse workers shoulder injuries. The arms are high-payload. The hands are dexterous. The whole point is manipulation.
And manipulation is where Dexmate has actually done the hard, unglamorous work. The company reports its dexterous hands achieve something like a 99% success rate across thousands of different objects. That number is the entire game. A robot that grasps correctly 90% of the time is a liability; a robot that grasps correctly 99% of the time is an employee. Getting from one to the other is not a demo - it is a slog through data, and Dexmate's approach is a "data flywheel" that blends massive simulation with real-world grasping data, refining the model with every attempt.
There is also a philosophy underneath the product roadmap. Dexmate talks about "Robot Copilot" before "Robot Autopilot" - meaning the near-term product ships with human teleoperation baked in, so a person can guide the robot through the tasks it cannot yet do alone. Every guided task becomes training data. It is a pragmatic admission that full autonomy is not here yet, dressed up as a strategy, which is the correct way to run a robotics company in 2026.
What is genuinely impressive, and slightly suspicious until you look closer, is the speed. Dexmate reportedly took Vega from company formation to a shipping, preorderable product in about six months. In hardware, six months is a rounding error. That they did it - and did it well enough that NVIDIA welcomed a barbell-lifting demo at its Santa Clara headquarters, and LG Technology Ventures wrote a check - suggests the team's research background translated into unusually fast execution.
Vega is priced at $89,999, with a $999 preorder deposit and lead times reported under four months. That is not a consumer product; it is a piece of capital equipment aimed at logistics, manufacturing, retail, and the growing crowd of researchers who want a physical platform to run their "physical AI" experiments on. Whether the world buys enough of them is the open question. But Dexmate has done the thing most robotics startups forget to do: it made a robot that is useful before it is impressive.
— Reporting compiled from public sources. Specifications and funding figures are approximate where noted.
