Moqi's Agentic-Native robot brain runs a 15-minute chore loop
Two Chinese embodied-AI startups just stepped out of stealth at the World Robot Conference with architectures that try to solve the field's hardest problem: not making a robot do one trick, but keeping it working when the task — and the environment — won't sit still.
Moqi Intelligent debuted an "Agentic-Native" embodied brain that ran a continuous 15-minute home-chore sequence with no human prompting. The company, founded just six months ago by former Huawei autonomous-driving AI lead Huang Qingqiu — a member of Huawei's "genius youth" program — showed its MORPHI KINO wheeled robot clearing a table, checking the fridge, fetching a water bottle to restock it, and running a wash-dry-fold laundry loop, all from a single plain-language goal like "tidy this room." The bet is that the long-horizon skill other labs bolt onto a big model as a separate planner should instead live inside the action model itself. Moqi's MoRA keeps the goal, a multi-layer memory of what it has done and what is left, and progress monitoring native to the policy network — so the robot self-corrects mid-task instead of freezing and asking the "brain" what to do next. It is a familiar complaint about today's System 1 / System 2 stacks, but the demo is unusually long for a home setting, and the company claims a valuation north of 7 billion yuan (about $1 billion) on the strength of it.
Lingxi Zhiyong, a new industrial embodied-AI startup, is attacking the same gap from the factory side with a "model + Harness" architecture. Founded in Shanghai by 1998-born USTC robotics PhD Duan Yifan, with his doctoral advisor Professor Ji Jianmin aboard as co-founder and chief scientist, the company pairs its CONWAY industrial-native model with a system called ROSS Harness — named for control-theory pioneer W. Ross Ashby. Rather than treating the model as the whole product, the Harness wraps it in an execution layer that handles model abstraction, reusable "skills," agentic planning, layered safety monitoring, and a data flywheel that turns each real run into recoverable experience. The pitch is explicitly anti-"model-only": pretrained models still succeed less than 60% of the time on real production lines, so Lingxi is selling the plumbing that makes a model into a shippable, recoverable system. ROSS Harness is set to appear at the second World Humanoid Robot Games.
The throughline is that embodied AI's center of gravity is shifting from "can the model do the task once" to "can the system keep doing it." We saw a similar push at WRC earlier this week — Daxiao Robot unveils Kaiwu World Model 3.1 at WRC — and the convergence suggests "long-horizon reliability" is now the real contest, not raw model capability.
Which matters more for robots leaving the lab — a smarter model or better execution infrastructure? Tell us in the comments.
