Faster than Bolt, but can China’s robots fold clothes?

(Edited and refined by Candice Chan, with the assistance of AI translation.)

The second World Humanoid Robot Games (WHRG) opened on 22 August at China’s National Speed Skating Oval, known as the “Ice Ribbon”.

In the 100-metre heats held before the opening ceremony, a robot athlete from Team Tianzhuo clocked 9.39 seconds, breaking the previous Games record, and surpassing the human 100-metre world record of 9.58 seconds held by Usain Bolt. The Tiangong Ultra humanoid robot also recorded a standing jump of 2.88 metres, exceeding both the human men’s high jump world record of 2.45 metres and the limits of the human standing high jump, in which both feet take off from a stationary position.

Humanoid robots breaking human sporting records is nothing new. At the Beijing E-Town Humanoid Robot Half Marathon held on 19 April this year, a robot named Lightning won in 50 minutes and 26 seconds, nearly seven minutes faster than the human men’s half-marathon world record of 57 minutes and 20 seconds set by Jacob Kiplimo.

But rather than humanoid robots repeatedly breaking human records at sporting events, what matters more is whether they are useful. Only when they are useful can they be deployed in the real world. And for that to happen, the industry needs to experience its breakthrough “ImageNet moment” and establish industry standards.

Less entertaining than humans

Looking at the most practical attribute of competitive sport — entertainment value — humanoid robots still lag far behind humans. Several other events at this year’s Robot Games showed that robot competitions remain far less compelling to watch than human ones.

In a football showcase, former Chinese national team player Yang Chen took a penalty kick that travelled at 85 kilometres per hour, while his opponent, a robot by Booster Robotics, managed a shot at 64 kilometres per hour. In this particular event, humans still have the advantage over robots in strength and speed.

Robots have even more shortcomings when subjected to the intense physical challenges of a real match. After a collision, for instance, they may struggle to get back on their feet quickly, making it considerably more entertaining to watch humans play football.

Another example was tennis, where China’s Grand Slam champion Zheng Jie took on a robot developed by Galbot. In the singles match, the 1.75-metre-tall robot demonstrated relatively good autonomous movement when faced with Zheng’s high-quality returns. It relied neither on human remote control nor pre-programming, instead making its own decisions entirely through visual algorithms and reinforcement learning, and was able to hit both forehand and backhand shots. Nevertheless, the robot lost every point. Zheng was essentially feeding it balls, and even then the robot sometimes staggered after making a shot.

It was much the same in table tennis, where Grand Slam winner and Olympic champion Ding Ning faced an Expedition A3 humanoid robot. While the robot could serve by itself and return shots on both the forehand and backhand, it still lost every point and missed several balls. And again, this was despite Ding feeding the robot balls on every point.

In all these respects, the robots performed worse than humans. They were particularly inferior in autonomy, instantaneous reaction time, speed, their ability to stop and accelerate abruptly, change direction and the quality of their shots.

Going beyond competitive results

Yet at this stage, the Robot Games are not about making robots move or behave like humans, either physically or ethically. Instead, the goal is to use real-world competition and practical tasks to expose problems, test capabilities and ultimately, establish industry standards.

There is an important term in the artificial intelligence (AI) and technology industries: the “ImageNet moment”. This describes the point at which a disruptive technology suddenly makes an explosive leap in performance, or the inflection point at which an industry moves almost overnight from laboratory research and development into large-scale commercialisation. The Robot Games are seeking precisely such a tipping point.

Countries around the world are developing AI and robotics, and companies everywhere claim that their robots are highly capable and impressive. Yet robots are being designed for different functions and tested under different conditions and in different environments, making meaningful comparisons difficult. In running, for example, some are tested on rubberised tracks, while others run on flat ground or even carpets. The Robot Games are like a standardised examination: robots are placed in the same setting and under the same conditions and allowed to put their unique skills to use, so that the best robots can then be identified and evaluated.

This could both produce an ImageNet moment and help establish industry standards, just as Formula One has enormous entertainment and commercial value not merely because it is exciting to watch, but also because F1 technologies have subsequently become benchmarks at the highest levels of the automotive industry.

In the robot 100-metre race, some competitors fell over, while others were unable to stop after crossing the finishing line and crashed into the barriers around the track. These incidents naturally provided moments of amusement, but more importantly, every fall generated data and revealed valuable “digitally expressible attributes or indicators”, or features. Such data provide the basis for subsequent improvement and evolution, and ultimately for an ImageNet moment and industry standards for robotics.

Practical, everyday functions

This year’s WHRG attracted 666 teams from 16 countries and 2,056 robots, four times the scale of last year’s event. The Games are therefore not merely about competing in competitive events like running, jumping and ball games, but also about competing in work and practical tasks, or scenario-based events. The number of scenario-based events has increased from six last year to 21 this year. Moreover, rather than taking place in purpose-built simulated environments, they are staged in real factories, real hotels and real rooms in homes, with the aim of testing robots’ ability to operate in actual living and working environments.

In the clothes-folding event, part of the household service and domestic-work competition, robots are required to complete a series of tasks involving putting away objects and organising clothes within a home environment in 30 minutes. A robot therefore has to make its own way to the bedroom, fold clothes, bedsheets and other items scattered across the bed, and then place them neatly inside a wardrobe.

In the cooking event, part of the food preparation and sales scenario competition, an actual food preparation and sales environment is recreated at the venue, testing the robots’ ability to carry out precise and delicate tasks with both hands. Robots have to operate various tools and tackle high-precision movements such as picking up food, measuring powders, using tweezers to pick up beans and opening bottle caps. These are crucial steps towards enabling humanoid robots to cook in kitchens and enter real-world industrial applications.

In the clothes-folding and storage competition, robots achieved a very high completion rate, but most were defeated by the washing and drying stage that followed. Taking clothes out of a washing machine and putting them on hangers requires highly sophisticated co-ordination between both arms; at the event, many robots had rigid grippers with just two or three fingers on one hand, making it difficult to secure a firm grip on the narrow curves of clothes hangers. When the other hand tried to put the garment onto the hanger, imprecise force control or visual parallax errors frequently caused the clothes to slip off or the hanger to fly off. Speed was another problem: a human can fold a garment in seconds, while a robot may take one or two minutes to do the same.

Generating useful data

Nevertheless, the data generated by all these actions will also provide an important reference for the future development of industry standards for domestic-service robots.

In the cooking competition, food-service robots have to clear three stages within 20 minutes. The first stage involves taking orders and collecting food. The robot must follow the referee’s instructions — such as “take one steamed bun and one dumpling” — pick up the correct items with tongs, and put them down steadily. The second stage involves heating the food. The robot must move autonomously to a microwave oven, open the door, put the food inside, close the door and operate the microwave to heat it. The third stage involves preparing a drink. The robot must pick up an empty cup, take it to the drinks dispenser and fill it with at least 200 millilitres of beverage before finally delivering all the food and drink safely to the designated collection area.

During the competition, tongs proved to be the undoing of many robots. Tongs are smooth, rigid and awkwardly shaped; a robot has to use binocular vision to identify them, apply precisely the right amount of force to grip them, and then use them to pick up soft steamed buns and dumplings. This is an extremely difficult test of “hand-eye-brain coordination” and immediately eliminated the less mature competitors.

All this data will likewise serve as a reference for future improvements to robots and the formulation of standards for domestic-service robots. Even as the Games are taking place, teams such as GigaAI in Wuhan, China, have already begun sending the first batch of general-purpose humanoid robots that have completed this kind of training — such as the SeeLight S1 — into real homes for large-scale testing. In actual household trials, the most advanced robots can already competently perform tasks such as heating chicken in a microwave, clearing plates and loading bowls and cutlery into a dishwasher.

Since what the Humanoid Robot Games are ultimately competing for is an ImageNet moment and industry standards, then for humanoid robots at the Games, after winning medals comes manufacturing orders, and going from the competition arena into the real world.

Related: China’s AI-driven robotics boom faces IPO reality check | The Buddha’s lesson for robots: Who’s real, who’s not?