At China's robot school, machines learn like humans, teacher uses VR headset

Yahoo

2 days ago

A specialized training facility in China has taken on its first cohort of 'robot students'. Located in Hefei—capital of the eastern province of Anhui—this new 'robot school' is being used to teach robots technical skills, such as holding and using tools.
The robot students are taught fine motor skills by a human teacher a wearing virtual reality (VR) headset and holding motion-sensing controllers. According to reports, one student robot arm is learning how to pick up a wrench and tighten a screw using this method.
The school, technically referred to as an embodied intelligent robot training environment, is essentially a high-tech training ground for robots, situated on a simulated factory floor.
"Our aim is to train robots to generalize from one example to others, so that they can perform reliably no matter the setting—and that only comes through real-world exposure," said Ji Chao, founder of a robotics company based in Hefei and a developer involved with the training facility.
"The service-based support model allows robot developers to purchase professional services regarding computing power, training scenarios, and data acquisition for algorithm iteration at reasonable costs, thus creating a virtuous circle," observed Sun Dandan, an executive at the robotics division of China's International Advanced Technology Application Promotion Center, which is independently running the program.
At present, robots from various Chinese companies are being taught how to function in specific work scenarios. These include, but not limited to, logistics and warehouse handling, picking and sorting parts, home assistance, and providing retail and tour guide customer service.
The training curriculum is pretty intense at the school, with human trainers inputting 200 action sequences per day per robot. These actions are collected as real-world physical data (not just code or simulations).
https://www.youtube.com/watch?v=-nTHdVlnnlU&pp=ygUYdGVhY2hpbmcgcm9ib3RzIHVzaW5nIHZy
Each robot uses this data to train machine learning models, which will eventually enable it to perform the task autonomously. The idea is not just memorizing a motion (e.g., turning a screw) but understanding how to adapt on the fly, such as recognizing different types of screws.
Real-world training like this exposes robots to unpredictable scenarios such as uneven surfaces, dropped tools, and unusual objects, which simulations often struggle to replicate well. The training is believed to help enhance the robots' ability to generalize across environments, making them more effective in real-world applications.
The school is the first public robot training platform of its kind in China. It offers shared infrastructure like computing power, datasets, and realistic environments, which is rare and expensive for smaller companies to build alone.
The initiative also supports various business models, and companies can co-run, operate independently, or purchase training and data services. The program helps bridge the 'gap', or the disparity between simulated training and actual performance in the real world.
It also promotes collaboration, standardization, and scalability in China's growing robotics sector. The ultimate goal is to expedite the development of more capable, general-purpose autonomous robots.
To sum up, the robot school is something of a real-world machine learning lab for robots which would enable them to work more independently in factories, warehouses, homes, or stores eventually.