AGIBOT holds World Challenge 2026 to see how AI models perform on real tasks

AGIBOT Innovation Technology Co., based in Shanghai, held the AGIBOT World Challenge 2026 in Vienna alongside ICRA 2026, bringing together 526 research and enterprise teams from 27 countries. The competition focused on embodied AI, testing models in two tracks: Reasoning to Action (R2A) and World Model (WM). R2A evaluated task understanding, planning, and execution, while WM assessed AI’s ability to predict physical-world changes and interactions. The challenge adopted a benchmark-driven format combining online automated evaluation with an offline real-robot final using AGIBOT’s G2 humanoid robot. This approach emphasized robot stability, real-world adaptability, and long-horizon task reliability, aligning technical evaluation with practical deployment needs. Over 100 teams surpassed the official baseline, with participants including the Chinese Academy of Sciences, Tsinghua University, Alibaba, and Russia’s Sber Robotics Center. During the finals, PrismBot from vivo secured first place with 43.47 points, followed by Shanghai RoboParty’s RP-VLA and Russia’s GreenVLA. The competition also introduced a supermarket benchmark track in collaboration with Dexmal, focusing on end-to-end decision-making and whole-body control in realistic retail environments. This track included challenges like object drops and grasping failures to better reflect real-world complexity. AGIBOT’s EWMBench and Genie Sim Benchmark provided a standardized framework for automated testing, reproducible results, and consistent metrics. Teams trained models using the AGIBOT WORLD open-source dataset, covering language understanding, spatial reasoning, and zero-shot transfer. The event highlighted a shift in the industry toward closed-loop testing on real robots and tasks, moving beyond simulation-based evaluations.