Robbyant’s Embodied AI Challenge Draws More Than 1,800 Teams From 13 Countries

Robotics and Automation News reported Nov. 9, 2026, that applications for the qualifying round of the embodied AI challenge hosted by Robbyant closed Oct. 27, with more than 1,800 teams from 13 countries registered. The competition is based on the open-source LingBot-VLA 2.0, and the final is scheduled to be held Nov. 14-16 in Shanghai as a real-robot hackathon.

The report said the competition will assess generalization in simulation and adaptation to real robots, making it a venue for developers to compete at the stage of connecting models to physical machines. The number of registrations shows the breadth of participation, but the final’s task conditions and results will need to be examined separately to assess performance on real robots.

More than 1,800 registrations indicate the scale of participation in a competition for embodied AI developers. However, the number of registered teams is not a measure of achievement on real robots or of model performance. The competition’s evaluation areas span model training and fine-tuning, generalization in simulation and adaptation to real robots. Participation and how well teams perform tasks on physical machines therefore need to be considered separately. A convenience-store trial in which six AgiBot G2 humanoid robots operated for 24 hours, reported by Gizmodo Japan, covered robot testing in a real-world setting, as we reported. This competition is a contest in which developers evaluate models, and differs in nature from a report on continuous operation in the field. Taken together, the two cases show that competing to adapt models from simulation to real robots and verifying operation in real-world settings should be treated as separate evaluation criteria. The report said the competition is hosted by Robbyant in collaboration with ModelScope and Alibaba Cloud Tianchi. Its design, which uses a common foundation model and is open to companies, universities, individuals and others, can be understood as providing a venue for developers to test approaches within the same framework. The number of participating teams alone, however, cannot establish a model’s reproducibility or applicability across different robots. The robots used in the real-robot hackathon, task conditions, scoring method, and number of finalists and their results could not be confirmed from the article provided; these are the next points to check when assessing real-robot adaptation.