Humanoid Robots Beat Bolt, but Folding Laundry Is Much Harder
Last month at the World Humanoid Robot Games in Beijing, humanoid robots broke Usain Bolt's 100-meter record. Running, jumping and breakdancing have become striking spectacles, but their practical usefulness is far less clear. The Beijing event highlighted a paradox: robots can outperform humans in controlled athletic feats, yet struggle with everyday chores.
For a robot, beating Bolt in a race is in many ways easier than folding clothes or working on a factory assembly line. Robots are beginning to show progress on those useful tasks, but such work requires complex hand movements and a brain that understands what it is doing—not blind forward sprinting until hitting a wall, as seen at the robot games.
The pace of learning complex tasks is accelerating. A report released this week by Citrini Research says bystanders underestimate their risk because robots are learning complex tasks faster and faster. Citrini does not expect the robotics industry to experience a single ChatGPT moment; instead, inflection points will arrive from multiple sources at once. That theory reflects the wide variety of tasks robots are being trained on, from manipulation to locomotion.
Recent advances have led analysts, including Citrini, to raise expectations. Last month Generalist showed a robot—essentially two arms with grippers—trained with its GEN-1.5 AI model. It can learn simple tasks such as stacking cups or paper folding by watching short video demonstrations. Soon after, Skild released a video of a similar robot learning to flip pancakes through demonstration. Figure's May progress update showed more commercial utility, with its humanoid team sorting 250,000 packages. Chinese startup Spirit AI's wheeled robots have already begun operating in a CATL factory, and its co-founder expects robots could enter commercial environments within one to two years. These examples suggest robot learning is broadening beyond locomotion to manipulation and real-world logistics.
Even once humanoid robots can handle tasks that make them useful, they face another barrier to mainstream adoption: they are expensive. JPMorgan recently estimated that robots capable of accelerating supply chains cost about $120,000 each. Elon Musk's long-term price target for Tesla's Optimus is $20,000 to $30,000, but Chinese startups, especially Xpeng, may reach that level before Musk does. The price gap corresponds to different mass-production paths. A $120,000 cost reflects current supply-chain levels and production scale, while $20,000 to $30,000 requires automotive-industry-scale manufacturing. That gap matters because it determines which customers can afford adoption.
Automakers such as Xpeng and Chery are entering the humanoid robot race, and their advantage lies in transferring automotive supply-chain cost control to robots. Qiyuan Robotics has already cut the price of its Q1 and T1 to 19,999 yuan, about $2,800—an order of magnitude below Musk's target. Of course, low-priced products have limited capabilities: the Qiyuan Q1 is 88 cm tall and weighs 15 kg, positioned for companionship and interaction, not heavy factory work. The transition from running to working depends on hand manipulation, brain understanding, and pushing costs down to a level customers will pay. The remaining challenge is closing the gap between athletic stunts and reliable, affordable work.