Field notes from IROS 2026 in Pittsburgh: a show floor full of dexterous hands, hand makers turning into data companies, and site-unseen demos that held up. https://x.com/i/article/2107581562409684992
IROS 2026 Recap
Boston Dynamics didn't have a booth at IROS this year. It didn't need one. For four days, close to twenty hand companies fought for floor space in Pittsburgh. Then, on the last day of the conference, @BostonDynamics put out a video of a new hand for Atlas and took the week anyway. It's the headliner move: skip the festival, drop the album the night it ends, and leave everyone on the side stages comparing themselves to you.
IROS 2026 ran September 27 to October 1 at the David Lawrence Convention Center, with roughly 4,000 attendees, close to 2,000 papers, and 183 exhibitors. It came four months after ICRA in Vienna, and the crowd was different again. Vienna was European, Hangzhou last year was Asian, and Pittsburgh was students, lots of them. Carnegie Mellon is a few blocks away and helped host, and it showed. "Student" undersells that room, though. A lot of them were presenting at the poster sessions, shipping hardware, and building companies already.
Everyone brought a hand
At least 17 dexterous hand companies exhibited, and 11 of them were Chinese. The list included Sharpa, WUJI, DH-Robotics, Linkerbot, PaXini, DexRobot, Xynova (its first US showing), BrainCo, Inspire, ROBOTERA, TESOLLO, Allegro, PSYONIC, and Chestnut. Another twenty or so were doing tactile sensing, force sensing, or data gloves. If ICRA's theme was "no new hardware," IROS answered with new hands. @chris_j_paxton called dexterity the biggest trend of the show, and he's right.
@SharpaRobotics had the biggest presence, with a new hand, an exoskeleton glove, a humanoid, and a stick-catching demo that ended up all over X. DH-Robotics launched a 13-DOF direct-drive hand on September 30. WUJI was the hand people kept bringing up, and it's reportedly shipping 800 to 900 a month. Most of the floor was competing on the same spec: five fingers, and 20 or more degrees of freedom.
Boston Dynamics went the other way. The new Atlas hand has four fingers and 13 degrees of freedom, up from 7 on the previous gripper. The actuators are fully enclosed, with no cables crossing the joints, and the whole thing was built for sim-to-real transfer and RL. The demos show it running drills, torque drivers, grinders, and nail guns.
The best part is how they decided to drop the pinky. The CTO had the team tape their pinky and ring finger together for a day to see what they'd actually lose. The verdict was no pinky. Four fingers and 13 DOF already covered what the hand needed to do, and a fifth finger would've meant three more actuators, plus the cost, volume, and failure risk that come with them. Their line for it is "the best part is no part."
That's the right way to build a robot. A hand doesn't get better by looking more human. It gets better by doing the job with the fewest parts that can break. Every extra degree of freedom is another motor, another encoder to zero, another thing to calibrate, and more action space for the policy to learn. It's telling that the most famous robotics company in the West landed on fewer fingers and fewer DOF in the same week a whole convention hall was competing on more.
Hand makers are turning into data companies
The more interesting shift was what the hand companies were selling next to the hands. WUJI, PaXini, and Chestnut were all showing UMI-style capture gloves, and Sharpa had its exoskeleton. @michellelsun laid out why. A lab buys a few hands, and then it needs far more demonstrations on those hands than it can collect itself. The hand's the easy sale. The data's the hard part, and the hardware vendors have figured out that whoever helps with the data keeps the customer.
The data vendors noticed too. At IROS Hangzhou last year, the only pure-play data company on the floor was Lightwheel. At ICRA, Encord said out loud that it does robotics. In Pittsburgh there were about a dozen, including Abaka, Datatang, iMerit, TELUS, Ropedia, and Lightwheel. Data's a booth category now.
The conversations on the floor moved with it. Fewer people asked how many hours you have, and more asked whether the data was designed for the task they actually care about. Volume gets you a model that can do the easy version of a task. It won't get you the version that needs precision.
The thousandth try
At ICRA I said what matters now is whose robot still works on the thousandth try, in a room nobody set up for it in advance. IROS moved in that direction, and you could see it in the demos.
The one that impressed me most was @LumosRobotics' towel folding. Look at those bimanual motions: almost human speed, and very smooth. There are a lot of companies I don't think could reach that level of performance. Like ours, it was a site-unseen demo. Nobody got to train in that hall before the doors opened.
Our booth ran a π0.5 checkpoint live on two YAM arms, picking fruit and dropping it into containers. The model was trained on data from a white-background work cell and deployed zero-shot onto a different rig. It ran in a convention center we'd never seen, with people walking behind the robot all day, and when we set up, we eyeballed where the arms went.
The usual lore says that shouldn't work. VLAs are fickle, they need consistent lighting, and you have to train on the exact machine you deploy on. Most of that turns out to be avoidable mistakes. Before the show we ran over 10,000 ablations in simulation, across lighting, backgrounds, camera placement, mounting, and training schemes. A handful of things mattered a lot:
- Fine-tune the whole model, not just the action expert.
- Use relative actions, so the policy tolerates small differences between rigs.
- Get the mounting height right. A 1 cm error cut success rates by something like 70%.
- Get the latency down.
Most other things mattered a lot less than people think.
The best reactions came from people who'd run a trade show demo themselves. They know how hard it is to get a learned policy running smoothly in a room it's never seen, and they asked the right questions. I'll write up exactly how we built it next week, and we're open-sourcing the code alongside it.
Other points of interest
- Universal Robots, the biggest winner at ICRA, didn't exhibit.
- Western venture-backed hardware came back. Agility, Dexmate, Standard Bots, and PSYONIC all showed up, after skipping Vienna entirely.
- A new FCC rule on foreign-made mobile robots may explain why at least one Chinese humanoid was shown on a fixed base.
- The most viral clip of the week was a Unitree G1 kicking a drone. Make of that what you will.
Zooming out
Put it together and IROS 2026 was the conference where the hand market got crowded and the data market got real. Those are the same story. Hardware's easy to buy now, and getting easier. What it learns to do in a room nobody prepared for comes down to the data behind it and how carefully you deploy it.
Next up is CoRL, November 9 to 12 in Austin. It's the most learning-heavy crowd of the year, which makes it the right room to test whether a demo holds up. We'll be there with a new one running. Come find us.