MindOne puts an AI housekeeper brain into Unitree's G1

Most humanoid robot demos come with an asterisk. Either the footage is sped up four times, or there is an engineer lurking just off-camera wearing a VR headset and mocap gloves.

A new video from a six-month-old AI startup out of Shenzhen is calling everyone's bluff with four bold words: no speed up, no teleoperation.

Crawling on beds and tossing frisbees

The company behind the clip is MindOne Robotics (Shenzhen Lingqi Wanwu Technology). In the video, a standard Unitree G1 humanoid navigates a household powered entirely by MindOne's proprietary "MindOn" AI brain.

The tasks shown are far beyond basic walking. The G1 is seen opening curtains, watering potted plants, collecting a parcel at the door, and wiping down a counter. In one of the most physically complex clips, it drops to its knees on a mattress to clean the surface, and later crouches down to reach trash wedged under a low table.

Toward the end, the robot is shown on an outdoor lawn playing a game of frisbee with children. Robotics AI expert Chris Paxton noted that seeing a robot interact with its environment at realistic human speeds, rather than moving in agonizing slow motion, was genuinely impressive.

Naturally, the internet met the video with a blend of awe and disbelief. Several developers on X pointed out that if these fluid, multi-step actions are truly running autonomously on an onboard model, it represents an unprecedented leap in edge robotics.

Supplying the brain, skipping the body

The skepticism is understandable because autonomous generalization remains the absolute bottleneck in humanoid robotics.

Unitree CEO Wang Xingxing candidly pointed this out recently at the Hongqiao Forum. Wang described the field's "ChatGPT moment" as reaching an "80/80" threshold: successfully handling 80% of tasks in 80% of completely unfamiliar settings. He admitted that progress across the entire industry has been slower than hoped, precisely because real-world training data is hard to come by.

MindOne is attacking the problem purely as a software player. Registered in May 2025, the startup just closed an angel funding round backed by Oriza Seed, Inno Angel Fund, and MSA Venture Fund.

Instead of building custom hardware, MindOne dropped its model into Unitree's G1. While teams at Stanford, MIT, and Amazon also use the G1, their published work usually focuses on isolated problems like fine pinch grips or safety balance. MindOne is the first to show a seemingly end-to-end, multi-chore household workflow on this body.

Why it matters

For anyone waiting for a robot butler, this demo shifts the focus from vanity tricks like backflips to real labor. Kneeling on a soft mattress and digging trash out from under furniture require complex contact dynamics, which are notoriously difficult for neural policies to master without losing balance. But do not rush to place an order yet; curated video clips in controlled environments rarely reflect the unpredictable chaos of a typical family home.

In our view, MindOne's strategy is undeniably smart. By avoiding the brutal capital expenditure of manufacturing metal and actuators, the company can piggyback on Unitree's mass-produced hardware. If its AI brain can reliably turn off-the-shelf bots into autonomous helpers, it positions itself as the default operating system for consumer humanoids.

The looming risk, of course, is verification. In a space burned by faked autonomy and quiet teleoperation, claiming zero human intervention sets an extraordinarily high bar. If MindOne cannot prove these behaviors work just as seamlessly in unseen, messy environments, the skepticism will quickly overshadow the hype.