ICRA 2026: The Humanoid Hype Met a Joystick in Vienna

The world's biggest robotics conference just wrapped in Vienna, and if you only read the press releases, you'd think the household humanoid had finally arrived. The reality on the show floor was more interesting — and more honest. Walking the ICRA 2026 exhibition floor, it became obvious that a lot of the humanoid robots on display were still being controlled by a human with a joystick. Not learning. Not reasoning. Being puppeteered. That single detail tells you almost everything about where humanoids really are in 2026.

The IEEE International Conference on Robotics and Automation (ICRA) ran June 1–5, 2026 at Messe Wien under the theme "Robots for All." It pulled in a record 5,088 paper submissions and more than 8,000 delegates, with over 140 companies and research institutes showing off their latest hardware. Humanoids were everywhere. So were the wires, the operators, and the joysticks.

The reality check nobody put on a banner

Here's the detail that tells you where humanoids actually are in 2026. On the exhibition floor, a lot of the most impressive-looking humanoids were not thinking for themselves at all — they were being driven, in real time, by a person holding a controller just out of frame. That's not a knock on the field. It's the single most useful thing you could know if you're trying to separate marketing from machine.

The consensus among the experts in Vienna was blunt: humanoids still need a great deal more AI training before they're genuinely autonomous in messy, real environments. The hard problem isn't making a robot that can walk across a stage — that's largely solved. It's reliable manipulation in cluttered, unpredictable spaces — the stuff that fills a real living room or a real warehouse — and that challenge surfaced again and again across the technical sessions. It is also, not coincidentally, exactly the capability a companion or home robot would need most.

"When you see a slick humanoid demo, the question to ask isn't 'can it move like a person?' It's 'who or what is driving it right now?'"

If you want a feel for the gap between the highlight reel and the show floor, here's our own clip from the conference beat — tap to play it (it loads from TikTok only when you do):

See it for yourself The humanoid reality check, in one short clip.
Tap to play. @robottesters on TikTok

What's actually moving the needle

The genuine momentum at ICRA wasn't a single robot — it was the software stack underneath all of them. Hugging Face's open-source LeRobot framework, integrated with NVIDIA's Isaac open models since January, has become the dominant robot-learning platform, with more than 58,000 community-contributed datasets now on the Hugging Face Hub. That's the kind of shared foundation that lets the whole field move faster instead of every lab reinventing the basics.

One caveat worth flagging if you ever plan to run it yourself: a critical remote-code-execution vulnerability (CVE-2026-25874) was still unpatched in the stable release at conference time. Great framework — but real security homework before you put it on a network.

On the research side, the recurring theme was robots that reason about the physical world more like we do — planning motions that include deliberate, controlled contact with their surroundings rather than treating every touch as a failure. That shift, from "don't bump into anything" to "use contact on purpose," is a quietly big deal for any robot expected to do real work with its hands, and it feeds straight into the AI brains driving these machines.

China set the backdrop

You can't talk humanoids in 2026 without talking about China, and ICRA was no exception. Chinese firms accounted for roughly 85% of all humanoid robot shipments in 2025, and Hangzhou-based Unitree alone shipped more than 5,500 humanoid units — about a third of the entire planet's output — at margins closer to a luxury brand than a hardware company. (If you want the human story behind that, we profiled Wang Xingxing, the engineer who made Unitree's robots cheap.)

Timed right around the conference, NVIDIA announced it had picked Unitree's roughly six-foot H2 humanoid as the body for the first robotics system it's selling to university researchers, pairing it with NVIDIA's Jetson Thor and Blackwell hardware. When the chip giant of the AI era chooses a Chinese humanoid as its reference platform for the world's top robotics labs, that's a signal about where the supply-chain center of gravity sits — a theme we keep coming back to in why robots are the next global trend after AI.

The debut that fit the consumer-robot beat best was Dyno, the first humanoid from Vietnam's VinDynamics (a Vingroup company), shown at ICRA and Computex. It's being positioned as a household assistant and a security/surveillance robot, and it was previously piloted as a multilingual tour guide at a safari park — a reminder that the most realistic near-term humanoid jobs are still "talk to people and narrate," not "do my dishes."

The near-term jobs that actually hold up

Put the show floor together and a clear pattern emerges. The closer a job is to structured, social and supervised, the more credible a humanoid is at it today. The closer it gets to open-ended manipulation in a messy space, the more likely there's a human on a joystick somewhere. Here's how the realistic roles stack up right now.

None of this is a reason to dismiss the category. It's a reason to read it correctly. The same manufacturers shipping these first buyable humanoids are improving fast — but the order in which capabilities arrive is "narrate and patrol" long before "tidy the kitchen unattended."

What this means if you're watching companion and home robots

For anyone tracking robots meant to live alongside us, ICRA 2026 lands on a useful middle ground between hype and dismissal. The hardware is genuinely impressive and getting cheaper fast. The autonomy is not there yet for the unstructured chaos of a real home, and a lot of what looks autonomous on a show floor still has a human in the loop. The most credible early roles are structured, social and supervised — guides, greeters, repetitive industrial tasks — not freeform domestic help.

There's also a quieter consequence buried in all that teleoperation: if a remote operator can drive a robot, a remote operator can, in principle, see and hear through it. That makes transparency about who is in the loop — and what happens to the camera feed — a real consumer question, not a theoretical one. It's the same thread we pulled in the real laws of robotics: the rules that matter most are the ones a maker is willing to show you.

Roland Berger's consultants in Vienna floated up to 50 million humanoids in use by 2035 in an optimistic scenario. That future is plausible. But the road there runs straight through the boring, hard work of AI training that this conference made impossible to ignore.

How to read the next humanoid demo you see

Take one habit away from ICRA 2026. When the next jaw-dropping humanoid clip lands in your feed, don't ask whether it moves like a person. Ask three smaller questions instead: Who or what is driving it right now? Is this one continuous take or a cut highlight reel? And is it doing the task in a space it controls, or one it doesn't? The answers usually separate a genuine capability from a beautifully staged remote-control puppet.

"The joystick on the show floor isn't a failure — it's a progress bar."

That's the honest read on humanoids in 2026. The bodies are ready enough. The brains are catching up in public, dataset by dataset and demo by demo. And the moment the joysticks start disappearing from the show floor for real — that's the signal worth waiting for.

Frequently asked questions