Ant Lingbo's First Embodied Large Model Challenge Concludes: TARS Team Wins Championship After 48-Hour Real-Robot Competition
On November 16, the final results of Ant Lingbo's First Embodied Large Model Challenge were announced. The TARS team, composed of contestants from the Futian Laboratory, Northwestern Polytechnical University, Pudu Robotics, and Tianjin University, won first prize with 82.5 points. The 15 finalist teams used the open-source pre-trained model LingBot-VLA 2.0 as their base and, under unified hardware and computing power, completed three real-robot tasks: single-arm grasping of fruits and vegetables, single-arm grasping of cosmetics and placing them into a transparent box, and dual-arm sweeping of desktop trash. On the same day, Ant Lingbo held its first developer conference and released its developer ecosystem plan, intending to invite 20 most valuable experts and recruit 40 developer experts; CEO Zhu Xing stated that the industry has not yet reached the Scaling Law moment for embodied intelligence.
Why read this
AI score 58
The competition unified the base model, hardware, and tasks; the teams' trade-offs in transparent object grasping and dual-arm collaboration can serve as engineering references for embodied deployment.
Topics
Body · AI translation
On November 16, the final results of the first Ant Lingbo Embodied Large Model Challenge were announced. The "TARS" team, composed of five contestants from the Futian Laboratory, Northwestern Polytechnical University, Pudu Robotics, and Tianjin University, won the first prize; the "EVA" team, composed of three contestants from Huazhong University of Science and Technology, the Shenzhen Research Institute of Tsinghua University, and Hubei Yuanyi Intelligence, and the "SST" team, composed of three contestants from Tianjin University, Peking University, and the Institute of Automation of the Chinese Academy of Sciences, won the second prize; the "Vida," "FDU_SONIC," and "ai-robot" teams won the third prize; another 9 teams won the Excellence Award, and 3 teams won the Observer Group Special Award. The TARS team won the first prize in the first Ant Lingbo Embodied Large Model Challenge. 48 hours of real-machine competition: robots had to pass three challenges. On November 14, the 15 finalist teams gathered at the Shanghai Zhangjiang Science Hall. The contestants came from Beijing, Shanghai, Hangzhou, Shenzhen, Chengdu, Wuhan, and other places, and some contestants even returned from Vancouver specifically to compete. At 9 a.m., as soon as the competition problems were announced, the teams plunged into the workrooms and began a 48-hour real-machine hackathon. In the competition, all teams used the same open-source pretrained model LingBot-VLA 2.0 as the technical foundation and carried out real-machine challenges using unified robot hardware, inference computing power, and task environments. There were three competition problems: first, grasping fruits and vegetables with a single arm; second, grasping cosmetics with a single arm and placing them into a transparent box; and third, using both arms together to sweep desktop trash and then pour it into a storage basket. At the competition site, contestants operated robotic arms to complete the two-arm collaborative task of sweeping up trash. These actions are very common in daily life, but in the hands of a robot, every step must withstand scrutiny. From recognizing objects, grasping them firmly, and placing them accurately—especially interaction with transparent objects—to the two robotic arms cooperating with each other, the examiners comprehensively assessed the quality of action completion and the task success rate. The finals consisted of three rounds of real-machine evaluation, and each team's highest total score from one of the rounds was counted in the ranking. In the end, the "TARS" team won first place with 82.5 points. This is a young team with an average age of only 25, and the team name is taken from the loyal and humorous robot in the film Interstellar. Team leader Wu Haonan is a post-00s contestant about to graduate with a master's degree from Tianjin University. During the roadshow, he presented the team's technical thinking and strategy: by simplifying complex multi-object scenes into single-object grasping, they effectively avoided the perception bottleneck of the VLA model in cluttered environments; they chose small-scale, high-precision data recollection and a two-person collaborative collection mode to avoid systematic errors of robotic arms in the physical world and to prevent human operators' physiological comfort from affecting the naturalness and effectiveness of robot motion trajectories. Li Yonglu, associate professor at the School of Artificial Intelligence of 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." By unifying the foundation model, hardware platform, and practical tasks, the competition built a "controlled-variable experimental field" for university scholars, industry engineers, and independent geeks. Every boundary condition of a failed grasp, every set of precisely calibrated supplementary data, and every real-world test feedback after model tuning can reduce the trial-and-error cost for those who follow and push the robot's "general brain" into the next real physical scene. Ant Lingbo adheres to embodied nativity and releases a developer ecosystem plan. On the same day, Ant Lingbo also held its first Developer Conference, where globally renowned scholars and frontline industry practitioners shared the latest research and practice in embodied large models and discussed topics such as difficulties in real-machine deployment and how to jointly build the developer ecosystem. In the scientists' technical lecture session, Shen Yujun, Chief Scientist of Ant Lingbo, elaborated on Ant Lingbo's embodied-native technical philosophy starting from real needs. He said that a smart robot brain faces three challenges: inputs have errors, outputs must withstand the test of physical interaction, and tasks in the physical world will not wait for the model; being able to solve robot problems does not mean being suitable for robots. In response, Ant Lingbo designs model capabilities around physical operations and builds an embodied-native brain, enabling robots to think more clearly, feel more sensitively, understand more thoroughly, and work more efficiently. Ant Lingbo Chief Scientist Shen Yujun elaborates on Ant Lingbo's embodied-native brain. At the industry summit forum, Ant Lingbo CEO Zhu Xing, BAAI President Wang Zhongyuan, Li Yonglu, associate professor at the School of Artificial Intelligence of Shanghai Jiao Tong University, Zheng Suibing, founder and CEO of RealMan Intelligent, and Chen Nutan, CTO of the Longsheng Weirui Embodied Intelligent Robot Innovation Center, also discussed topics such as difficulties in real-machine deployment and how to jointly build the developer ecosystem, with Han Chenglong, head of embodied intelligence investment at Ant Group, serving as moderator. The conference also officially released the Ant Lingbo Developer Ecosystem Plan, which plans to invite 20 Ant Lingbo Most Valuable Professionals to support teaching and academic innovation and jointly cultivate embodied intelligence talent; at the same time, it will recruit 40 Ant Lingbo Developer Experts to support technical practice and professional growth and jointly build the developer ecosystem. For the burgeoning embodied intelligence industry, Zhu Xing, CEO of Ant Lingbo Technology, emphasized in an interview that the industry is still iterating rapidly and has not yet reached the Scaling Law moment for embodied intelligence. However, with the validation of some technologies in real-world deployment and collaboration across industry, academia, and research ecosystems, the evolution of embodied intelligence will also accelerate. The Ant Lingbo Embodied Large Model Challenge is an embodied intelligence algorithm challenge open to global enterprise technical experts, university research teams, individual developers, and technology enthusiasts. It has two stages: an online algorithm preliminary round and an offline 48-hour real-machine hackathon final. The event attracted registrations from 1,855 teams and 3,252 contestants from 13 countries around the world, covering 648 universities, research institutes, technology companies, and embodied intelligence teams, making it the largest embodied large model challenge in the industry.