AI Trends

Ant Lingbo's First Embodied Large Model Challenge Held in Shanghai: 15 Teams Compete in 48-Hour Real-Robot Grasping Contest

Shanghai ObserverUpdated Nov 16 21:21

AI summary

The finals of Ant Lingbo's First Embodied Large Model Challenge were held at the Shanghai Zhangjiang Science Hall. The 15 finalist teams, using the open-source pre-trained model LingBot-VLA 2.0 as their base, competed in a 48-hour real-robot challenge featuring three tasks: single-arm grasping of fruits and vegetables, single-arm grasping of cosmetics into boxes, and dual-arm sweeping of desktop trash. Ultimately, team "TARS" won the championship with 82.5 points. The competition attracted a total of 1,855 teams and 3,252 contestants from 13 countries. Ant Lingbo CEO Zhu Xing stated that embodied intelligence has not yet reached its Scaling Law moment, while Chief Scientist Shen Yujun predicted that the industry may see its first wave of convergence.

Body · AI translation

What can an embodied large model do in 48 hours? The finals of the first Ant Lingbo Embodied Large Model Challenge held in Shanghai posed an extreme test. At the Zhangjiang Science Hall, simple green-screen cubicles were set up in a row, and young developers were focused, either staring at screens and typing code or gathered around robots making adjustments. Fifteen teams that reached the finals pitched tents here and used 48 hours of extreme time to challenge the grasping capabilities of humanoid robots. Experts generally believe that embodied intelligence has not yet reached its Scaling Law moment, but may soon usher in its first wave of convergence. A 48-hour real-machine contest in which robots had to pass three tests. On November 14, the 15 finalist teams assembled at the Zhangjiang Science Hall in Shanghai. Contestants came from Beijing, Shanghai, Hangzhou, Shenzhen, Chengdu, Wuhan and other places, and the contestant who traveled the farthest came back specially from Vancouver to compete. At 9 a.m., as soon as the challenge problems were announced, the teams plunged into their workrooms and began a 48-hour real-machine hackathon. The competition was an assigned-task format: using the open-source pretrained model LingBot-VLA 2.0 as the technical foundation, teams carried out real-machine challenges with unified robot hardware, inference computing power and task environments. There were three challenge tasks: grasping fruits and vegetables with a single arm; grasping cosmetics with a single arm and placing them into a transparent box; and using both arms together to sweep desktop trash and then pour it into a storage basket. Jiang Bo, head of market operations and technical ecosystem at Ant Lingbo Technology, said that although these actions are very common in daily life, they are not easy for robots. From recognizing objects, grasping them firmly and placing them accurately to the two robotic arms cooperating with each other, the examiners comprehensively assess the quality of task completion and the success rate. In the end, the TARS team won first place with 82.5 points. This was a young team with an average age of only 25, and its name was taken from the loyal and humorous robot in the film Interstellar. Team leader Wu Haonan, a post-00s contestant about to graduate with a master's degree from Tianjin University, presented the team's technical thinking and strategy during the roadshow. By simplifying complex multi-object scenes into single-object grasping, the team effectively avoided the perception bottleneck of VLA models in cluttered environments. They chose a small-scale, high-precision data recollection approach and a two-person collaborative collection mode to avoid systematic errors of robotic arms in the physical world and to prevent human operators' physical comfort from affecting the naturalness and effectiveness of robot motion trajectories. The reporter noted that many finalist teams adopted the most cutting-edge technical paths. The EVA team, which won second place, was formed by a mix of students from Huazhong University of Science and Technology, Tsinghua University, Peking University and other institutions. Team leader Wu Xianjin said that when the team received the challenge problems, they thought of using the currently hottest RSI route to make the robot quickly get to work, and after more than 100 autonomous tests, they got the RSI flywheel spinning. Embodied intelligence is still iterating and may usher in its first wave of convergence. Data shows that the first Ant Lingbo Embodied Large Model Challenge set the largest scale in the industry, attracting a total of 1,855 teams and 3,252 contestants from home and abroad to register. Participants came from 13 countries and covered 648 well-known universities, research institutes, leading technology enterprises and embodied intelligence companies. The 48-hour real-machine contest appeared on the surface to be 15 teams grasping fruits and vegetables, collecting cosmetics and clearing desktops, but behind it was a stress test of embodied large models moving from screens to physical sites. In Jiang Bo's view, unifying foundation models, hardware carriers, inference computing power and task environments in a short period of time is very difficult, and participating teams need to quickly locate problems, find their root causes and then solve them. Judging from the finals site, some teams chose to recollect data, while others chose to work together with large models such as DeepSeek. Li Yonglu, associate professor at Shanghai Jiao Tong University, believes that what the Embodied Large Model Challenge has precipitated through competition is not only honors, but also a set of engineering experience and methodology that has withstood real-world testing: every boundary condition of a failed grasp, every set of precisely calibrated supplementary data, and every measured feedback after model tuning are lowering the trial-and-error cost for those who come later and pushing the robot general brain into the next real physical site. Embodied intelligence is still iterating rapidly and has not yet reached the Scaling Law moment. Ant Lingbo CEO Zhu Xing said bluntly that before the arrival of Scaling Law, it is first necessary to determine which kind of data scaling can promote the emergence of intelligence, but the industry has not yet reached consensus or an answer. Ant Lingbo Chief Scientist Shen Yujun pointed out that at present, none of the key links in embodied intelligence have converged. The lack of convergence in evaluation benchmarks and the difficulty of model convergence are mutually causal, and data collection forms are also not unified. But he said that with long-term exploration by the industry, consensus is forming: in the near future, the industry may usher in its first wave of convergence, involving evaluation mechanisms, model approaches and data collection standards. At that time, industry development will clearly accelerate, and role division and development directions will also become clearer.

Source

Shanghai Observer