Beijing opens "Kindergarten" for robots: Machines learn from their own mistakes

An unusual research centre—a "kindergarten" for robots—has opened in Beijing's Shijingshan district. Here, machines learn to move and interact with the world not through pre-set human instructions, but through trial and error. The project is a collaboration between Chinese company Tashan Technology and a team led by Turing Award laureate Richard Sutton, a pioneer of reinforcement learning.

What Robots Are Taught

According to globaltimes.cn, the conventional approach to training humanoid robots relies on human demonstrations, remote control, or motion capture. In China's new training centres, machines often practise household and industrial tasks in exactly this way. The "robot kindergarten" asks a different question: can robots continue learning after deployment, drawing on their own experience?

The centre is equipped with tactile sensing systems. Robots repeatedly perform movements, receive feedback from physical contact, and adjust their behaviour. If a machine bumps into a wall, the collision becomes part of its experience: the system records the error and tries to avoid it next time. Failure here is not a malfunction but training data.

Why This Matters

At the opening, Richard Sutton noted that the project requires a safe environment in which robots can explore the capabilities of their "bodies" and the laws of the physical world. This demands not only new algorithms but also specially designed machines and spaces. Mistakes are inevitable—both researchers and robots will have to learn from them.

Experts highlight the difference between short-term and long-term effects. Learning from human demonstrations yields quick results, but the robot's behaviour often remains fixed afterwards. Continuous learning through experience, in theory, allows a machine to adapt to new objects and situations not present in the original training data.

The weak point of this approach is the hardware. The robot must be robust enough to survive its own errors. If a machine cannot afford to be wrong, it cannot learn either.

Prospects and Limitations

According to specialists, self-directed exploration helps robots move beyond narrow, pre-defined tasks and develop a more general understanding of the physical environment. Over time, experience from individual machines could theoretically be transferred to shared models and propagated to other devices. This is especially valuable in complex environments where it is impossible to gather demonstrations for every scenario in advance.

For now, the method remains at an early stage. A small spider robot at the centre learned to move forward in about 40 minutes with no prior knowledge—but this is a narrow task. For complex humanoid robots, the variables are far greater: balance, energy consumption, temperature, and safety.

Most likely, learning through self-experience will complement, rather than entirely replace, teleoperation, imitation, and simulation. However, as robots move from laboratories to factories, services, and homes, the ability to adapt after deployment could become one of the industry's key development directions.

In Brief

A "robot kindergarten" has opened in Beijing, where machines learn through tactile experience, trial and error, rather than solely through human demonstrations. The project is run by Tashan Technology in collaboration with Richard Sutton's team. The approach allows robots to accumulate their own experience and, in theory, adapt to new situations, but it remains at an early stage and requires robust hardware capable of surviving errors. This method is expected to complement existing training approaches.

Follow NEWS.am Tech on Facebook and Twitter