Zhiyuan Robotics Just Dropped 4 New Robots—And This Time It Actually Feels Different

After attending Zhiyuan Robotics’ partner conference in Shanghai last week, the first thing I want to say is: this time felt different.

Not different in the “PPT demo” sense—different in the sense that actual hardware was there, you could touch it, and code was actually running on it.

On April 17th, Zhiyuan dropped: 4 new robot bodies, 4 AI foundation models, 7 solution packages, and an open dataset. 2,500 attendees, 34 countries—the scale itself signals something.

Which of the 4 new bodies is worth watching?

My personal interest landed on a detail most coverage missed: their new quadruped robot, the KuTuo D2 series, was shown in actual industrial deployment—grain warehouse inspection, submarine cable detection. Not lab demos. Real industrial jobs.

This tells me something important: embodied AI is graduating from “can walk and run” to “can actually do work.”

The humanoid half-marathon is exciting, but commercial deployment is way harder than running 21 kilometers. Walking 10K and completing an industrial inspection task are completely different challenges. Zhiyuan prioritizing “operational capability” over “performance showcase” seems like the right call to me.

Can LLMs and robots actually fuse deeply?

Everyone’s been saying “LLM + robot” for two years, but in practice, language models and control systems have been two separate stacks—language models think, control systems move, and there’s a massive gap between them.

What caught my attention in Zhiyuan’s approach this time: they’re trying to打通 LLM reasoning and robot motion control at the foundational level. Not simple “LLM outputs instructions, robot executes.” Rather, the robot can call LLM chain-of-thought capabilities during actual movement.

In plain terms: the robot isn’t just “following orders”—it can think while moving, autonomously reason through what to do next when something unexpected happens.

I don’t know if this will fully work. But directionally, it’s a lot more serious than “putting ChatGPT in a robot.”

My one reservation

After all the positive stuff, I have one concern: how much of this is actually production-ready versus partnership/fundraising theater?

The robotics industry has a new “world’s first” or “breakthrough” announcement every month. And most of what gets announced never runs reliably in a real factory for more than six months. How many of Zhiyuan’s new products are truly engineering-ready versus hype with a conference badge?

That said, the April 19th half-marathon was actually a decent real-world stress test. Flawed, messy, nobody finished perfectly—but publicly running 21 kilometers with a robot is a legitimate engineering validation.

The results were what you’d expect: some robots fell, some stopped, none were flawless.

And that’s actually the point. Embodied AI is still extremely early. There’s real work left to do.