AINEWS
Outside autos and electronics, the bottleneck isn't a better model but the unit economics of getting a robot to learn from the line it sits on.
Outside autos and electronics, fewer than one in five factory lines runs a robot. The reason is not a missing product.
The global industrial-robot fleet grows by roughly 500,000 arms a year, but most of them end up in car plants, electronics assembly, and a handful of other high-volume sectors. General manufacturing, the welding shops, machine-tending cells, paint lines, and small-batch assemblers that make up most of the world's factory floor, sits below 20% robot penetration according to Leiphone's review of the sector. The bottleneck has shifted from capability to deployment economics: getting a robot to learn from the line it already sits on, at a price the line can pay.
Four engineering tensions keep that gate closed. Training a new welding or painting worker takes one to three months, and turnover is high. Models trained in simulation break the moment the line changes SKU. Small batches break the ROI math. And the data flywheel that would make robots adaptable does not start spinning until real-line deployment demonstrates the system works. Each tension makes the others harder.
China-based industrial-robot maker Efort has placed a concrete bet on one answer to all four. In May 2024, the company spun out 启智(芜湖)智能机器人有限公司 (Qizhi), a joint entity with the National Advanced Manufacturing Industry Investment Fund and Wuhu Sci-tech Fund, to build what it calls a "dual-wheel drive" of self-developed hardware and an intelligent brain, per the Efort newsroom. Qizhi reported a near-100M yuan angel round (~$14M USD at approximate mid-2024 rates) led by Guotou Zhaoshang, with the funding receipt trail carried by Caifuhao Eastmoney.
The stack Efort and Qizhi have built around that bet is wide. The company describes an Openmind operating system, a Modou IDE, a Dayan data platform, a HumanGPT world model, and a HALO skills service it has labeled the "Android of robots," a phrase borrowed from the mobile-app era to signal a platform ambition. HALO is the data side of the bet: a wearable capture suit with 8K panoramic vision, depth sensing, fingertip and palm tactile arrays resolving below 0.01 newtons, surface EMG, and a 27-node IMU running under 10 ms latency. The data pipeline runs bottom-up, atomic skills first, then combinations, then task skills, designed to transfer across scenarios.
The neural model handles long-horizon planning, decision-making, and vision; classical algorithms guard collision detection, precision, and robustness on the line. Line-side bin-picking is the worked example the company points to. The open question is whether wearable capture output survives the trip from operator to model to a line that changes SKU every week.
Stcn coverage of the spin-out frames the capital bet as a vote that the last mile is solvable, not that it has been solved. The Leiphone piece is closely aligned with the Efort and Qizhi narrative, and product claims around Openmind OS, HALO specifications, and the HumanGPT world model should be read as company-sourced until independent benchmarks or customer lines confirm them. No outside customer, integrator, analyst, or competing supplier on the record in the current source set says the model-plus-rules hybrid and HALO capture actually close the gap.
The watch item is real-line deployment traction. The angel round buys Qizhi roughly twelve to eighteen months to convert HALO capture output into a line a paying customer will run for more than a pilot. If wearable data does not survive the trip to a high-mix line, the flywheel never starts, and the last-mile gap stays a sector-wide problem. If it does, the unit-economics gate, not the model, was the thing that had to fall.