Wuhan builds "Robotics Vocational College": customized AI teaching materials for robots

China News Service, Wuhan, October 28 (Reporter Liang Ting) In Wuhan's high-quality data base, trainers collect real operation data through lightweight intelligent equipment and customize exclusive AI teaching materials for embodied intelligent robots. Reporters learned on the 28th that since its launch in August this year, the "Robot Vocational College" has had more than 16 million downloads of its embodied data set globally, serving more than 80% of domestic embodied intelligence manufacturers.

In the restored supermarket scene, the trainer wears a 240-gram lightweight smart headband to complete operations such as product sorting, shelf sorting, and product replenishment. Mobile phones and lightweight portable devices record the interaction between hands and objects from the perspective of workers.

On October 28, at the Wuhan High-Quality Data Base, data collectors wore lightweight smart headbands to complete operations such as sorting goods, sorting shelves, and replenishing goods. Photo by Xu Liang

Zhu Kai, deputy director of Ant Digital Tianji Laboratory, said that 82% of the front-line collection personnel are original practitioners in the industry, relying on their job operation experience to produce original data that is real, practical and in line with industry needs. After the videos are collected at the Wuhan base, "what the person is doing and when the action occurs" is annotated, and then the movement and contact process of the hand is restored through trajectory reconstruction, forming data for the model to learn.

At present, the collection scenes are not limited to the internal training venues of the base, but have covered more than 1,100 real industrial scene collection points such as pharmacies, supermarkets, and laundry rooms.

Currently, the scale of Hubei's artificial intelligence industry continues to rise, and the layout of humanoid robots and embodied intelligence industry chains is accelerating. High-quality real-life training data is the key for AI to move from the laboratory to industrial scenarios.

On October 28, in the training area of ​​Wuhan High-Quality Data Base, robots were trained in simulated pharmacy scenarios. Photo by Xu Liang

Zhu Kai introduced that from "experience" to "available data", problems such as continuous action recognition, occlusion and perspective changes need to be solved. The team improves annotation quality through dedicated annotation models, multi-model collaboration and manual review. The system can accurately analyze camera motion trajectories, hand joint postures and object contact relationships. The hand joint position reconstruction error is only 7 mm. It also relies on a model adaptation library covering more than 2,000 mobile phone models to achieve data standardization and unification.

Data show that in 2025, the scale of Hubei's artificial intelligence industry will reach 160.8 billion yuan (the same below), with a compound growth rate of nearly 40% in the past five years. Up to now, many universities in Hubei have established robotics colleges, and more than 60 higher vocational colleges have established robotics-related majors. The scale of the province's embodied intelligence industry has exceeded 40 billion yuan, and there are more than 140 related companies in the humanoid robot industry chain.