This Chinese startup says teleoperation is poisoning humanoid robot training data
The dominant method for training humanoid robots to do physical tasks is teleoperation: a human operator controls the robot's limbs in real time while cameras record the movements. That data then trains the robot's policy. Most of the major players — including many backed by hundreds of millions of dollars — use this approach.
MindOn says it's fundamentally wrong.
"Human-form robot training is progressing slowly because the mainstream teleoperation-based approach has a fundamental flaw," co-founder Zhu Qingxu told KR Asia.
The flaw, as MindOn describes it: conscious, deliberate human control is slow and jerky by nature. When a person focuses on operating a robot's arm, they move differently than they do when performing the same task naturally. The resulting training data encodes artificially slow and stilted movement patterns. Robots trained on it inherit those characteristics.
MindOn's alternative is to skip teleoperation entirely and train exclusively on human motion capture data ... operators in mocap suits performing tasks naturally, with the resulting fluid, subconscious movement patterns fed directly into the model.
The speed claim
The company calls its AI model Mind-1, and as Interesting Engineering reports, the headline claim is human-speed task performance. A September 30th demonstration video was explicitly labeled "1.0x Speed — Fully Autonomous." The "1.0x" designation was a deliberate signal: no speed-up was applied to the footage. A subsequent logistics demo was labeled "1.5x Speed Fully Autonomous," meaning the robots were operating faster than a human baseline.
The demo showed Unitree G1 humanoids handling parcels, folding shirts, and collecting laundry. An earlier November 2025 video showed a G1 autonomously opening blinds, watering plants, delivering packages, ironing a shirt, sorting toys, and throwing a frisbee ... all at natural human pace, all without teleoperation.
Those are company-reported claims, not independently verified, mind you. Interesting Engineering noted that the footage provides no quantified human-versus-robot comparisons, task success rates, or intervention statistics.
MindOn has no named commercial customers yet.
The architecture
MindOn's model uses a two-layer architecture deliberately modeled on the human brain.
A "cerebellum" layer handles motion control: walking, running, crouching, grasping, pulling. A "cortex" layer handles perception, planning, and task reasoning, calling on the cerebellum's skills to execute. The company also uses what it calls a "Real-World Execution Compensation Model" to bridge the simulation-to-reality gap, which it claims achieves sub-1 cm manipulation accuracy on the Unitree G1.
Crucially, MindOn is a pure software company. They don't manufacture robots: Mind-1 runs on third-party hardware, primarily Unitree platforms. The model is designed to be cross-embodiment: the same underlying intelligence layer runs on both bipedal humanoids and stationary dual-arm systems, as demonstrated in a multi-robot logistics coordination demo showing two G1 units and two stationary arms collaborating on a single packing workflow.
The context
MindOn was founded in Shenzhen in May 2025 by Zhu Qingxu and Zhou Qinqin, both ex-Tencent Robotics X researchers. The company raised approximately $14 million across three rounds in its first six months, from investors including MSA Capital, Genspark, Oriza Seed, and Plum Ventures.
The companies it is challenging — Physical Intelligence (pi.ai), Skild AI, Covariant — are orders of magnitude larger and better-funded. Physical Intelligence alone has raised more than $400 million and has commercial deployments. Skild AI claims 1,000 times more training data than competitors. These aren't startups; they're shaping the standard approach.
MindOn's bet is that the standard approach has a structural deficiency that more data and more funding won't fix. The argument is methodological, not just competitive: if teleoperation training produces inherently slow data, scaling it up produces more of the same slow data. The ceiling on the technique is low.
Whether that thesis is correct is genuinely unknown. Robot AI is a field where claims routinely exceed reproducible results. But the anti-teleoperation argument is coherent, the demos are specific about speed (which most robot demos are not), and the team has direct experience from one of China's most serious robotics research programs.
Zhu is direct about the timeline: "I believe the right robot form factor for real household scenarios is still a bipedal humanoid, and it should be achievable within three to five years." Whether MindOn gets to prove it at commercial scale before better-funded competitors either adopt the same methodology or prove it wrong is the interesting question.
Sources
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”