Sprint before the Humanoid Robot Games to prepare for the "Housekeeping Competition": Complete the storage and folding of clothes in 30 minutes, and the large model helps continuous learning

IT House reported on August 16 that the 2nd World Humanoid Robot Games will kick off on the evening of August 22 at the "Ice Ribbon" of the National Speed ​​Skating Stadium in Beijing. As the competition period approaches, each participating team has entered the final sprint training stage. Starting today (16th), preliminaries have started for some scene competitions. Therefore, CCTV News visited the pre-match training situation of some teams today. This Games is co-sponsored by the Beijing Municipal People's Government, China Central Radio and Television, etc. The main competition will be held from August 22 to 26, a total of 5 days, and a total of 1,301 games in 51 events will be held. The event attracted 666 teams from 16 countries on six continents, with a total of 2,056 robots participating. The number of teams increased by 138% compared with the first edition, and the number of robots quadrupled. At the humanoid robot industry ecological training and evaluation base at the Beijing National Speed ​​Skating Arena, participating teams are training around nine real-life scenarios including homes, hotels, and industries. In housekeeping service scenarios, robots need to complete multiple tasks such as storing items and folding clothes, which places extremely high demands on the robot's hand-eye-brain coordination, long-term continuous work, and multi-task autonomous planning capabilities. In order to teach robots to handle flexible materials such as clothes and sheets, one team equipped the robot's head and hands with high-definition motion cameras. Data collection personnel and robots cooperate with each other to use the collected scene images, item parameters and action data for large model training to continuously iteratively optimize the system. Zhang Zhizheng, co-founder of a robotics company, said that the team constructed five different types of data, including Internet data, ontology-free human data, cross-ontology simulation synthetic data, remote control operation data, and real test environment closed-loop return data, which play a complementary role in model training. In response to the wear and tear problem of robot joint parts after long-term use, the team also developed a "lifelong learning mechanism" to allow the robot to continuously adjust its motion program during work and maintain long-term and stable task execution capabilities. According to the requirements of the competition, the robot must complete multiple tasks such as storing items and organizing clothes within 30 minutes. In the meantime, come