China is training thousands of robots for the real world

China's humanoid robot industry is expanding rapidly on the hardware front, but a shortage of data and the high cost of software development are emerging as major obstacles to making the machines useful in everyday life.

Sales of humanoid robots in China are projected to triple to 50,000 units this year, while manufacturers have made advances in physical capabilities. Unitree's humanoid robot, for example, briefly reached a running speed of 12.66 metres per second, described as a "Superman" milestone, says The Economist.

But improving robots' ability to perform complex tasks requires major advances in software. Developers need to scale foundation models from a few billion parameters to hundreds of billions, requiring vast amounts of real-world physical data.

Such data includes information that is difficult to capture digitally, such as spatial awareness, the brittleness of objects and the viscosity of liquids. Collecting it is considerably more difficult than gathering the text and other digital material used to train AI chatbots.

China builds a data-gathering network

To address the shortage, China has launched a large, state-supported effort to collect physical data for humanoid robots, using two main approaches.

The first involves what developers call "real machine" data, generated when robots perform tasks either autonomously or through remote control. China already has 53 specialised training centres, with another 34 planned or under construction, where robots spend hours carrying out tasks modelled on human activities, including stocking shelves, working on factory lines and pouring coffee.

The second approach, known as "egocentric" data, records human movements using sensor-equipped equipment such as haptic-sensor gloves and headsets.

A growing group of wearable-equipment suppliers has emerged around the industry. Companies including JD.com are paying hundreds of thousands of employees and part-time workers to record their movements in simulated environments, generating data that can be used to train robots.

The Chinese government is providing substantial financial support for the effort, covering about 80% of the cost of training centres.

That backing is important because collecting physical training data is expensive. The process can cost between 500 yuan and 700 yuan per hour, largely because much of the data generated during training is unusable due to errors that engineers do not want robots to learn.

Local governments are also working with private companies, purchasing robots to support local manufacturing and then selling the resulting training data to companies.

A different approach in the US

US companies are pursuing less capital-intensive ways of gathering training data.

Tesla is reportedly training robots directly in its warehouses, while other Western companies are turning to video-based training or employing lower-wage workers in developing countries to wear equipment used to collect egocentric data.

China's approach is more labour-intensive and expensive, but its scale could give the country a significant advantage in collecting the physical data needed to train humanoid robots.

Even so, a commercially viable mass market for humanoid robots is unlikely to emerge this decade, according to the outlook described.

In the near term, Chinese companies are expected to produce hundreds of thousands of expensive machines with limited functionality. The longer-term objective is to use the resulting data and experience to develop cheaper robots capable of performing useful tasks as everyday assistants.