Millimeters, not megawatts: what really holds back humanoid robots

Talk about robotics usually starts with compute power and degrees of freedom. The bottleneck can be far more mundane. At the fifth Chinese digital trade fair in Hangzhou on 24 September, Unitree founder Wang Xingxing argued that precision is the core problem. An AI model does not match its inputs and outputs to physical reality, he said, so robots work with errors on the order of a few millimeters. Whoever solves that problem solves the problem of robots.

Wang also set out when he would call the industry mature. His test: a robot completes 80 percent of tasks in 80 percent of scenarios it has never seen before, using speech and its own body. That, he says, is the "ChatGPT" moment for embodied artificial intelligence. For now, as he admits, productivity is what keeps humanoids out of factories and homes. The robots can do things, but more slowly than a person, and new tasks require retraining.

A market that is already moving

The barriers do not mean stagnation. According to the Chinese outlet IT之家, citing Counterpoint Research, around 35,000 quadruped robots were sold worldwide in the first half of 2026, three quarters of the full-year figure for 2025. Industrial applications took 40 percent of that number. Chinese companies dominate the market: Unitree, AGIQUAD, Genisom AI and DEEP Robotics. Genisom delivered more than a thousand units of its heavy "铜锤" series. Analysts forecast that sales will exceed 90,000 units in 2026 and revenue will double to 1.5 billion dollars. By 2030, they say, revenue could pass 10 billion dollars, with average annual growth above 60 percent. The most promising segment, in their view, is machines with a dynamic payload above 20 kilograms.

The division of labour is telling. Unitree aims mainly at entertainment, education and research, while AGIQUAD and DEEP Robotics focus on industrial applications.

The other side of the coin

Set that data against research into what robots find hardest: precision of movement. The German outlet t3n describes the tiny DirectHop robot from the University of Washington. It is 50 millimeters wide, weighs one gram and can jump a typical stair step. An electric motor and three adjustable legs let it launch itself to a set height with an accuracy of one centimeter. Earlier flea-inspired designs struggled precisely with jump calibration: springs and latches made adjustment difficult. The robot will be presented on 30 September at the IROS conference.

In robotics, progress today is measured not in model parameters but in the repeatability of movement in an unpredictable environment.

Put the two threads together and the conclusion for investors and industrial buyers is sober. Chinese companies have mastered the sale of machines that already earn money: inspection, logistics, site patrols. Humanoids, which the capital market is waiting for, still require solving the problem of millimeters, and that is research work, not production work. The DirectHop story points the way. Meaningful progress is not another acrobatic trick, but a robot that precisely does exactly what it is told.