AGIBOT hosted the AGIBOT World Challenge 2026 alongside ICRA 2026 in Vienna, with 526 research and enterprise teams from 27 countries competing in two embodied AI tracks focused on reasoning, action and world modeling. The competition was structured around two evaluation areas. The Reasoning to Action track assessed how robots interpret tasks, plan actions and execute them in physical environments. The World Model track evaluated how AI systems predict changes in the physical world and model interactions based on robot actions and sensor inputs.
Participating teams came from universities, research institutes, technology companies, startups and individual developer groups. Institutions and companies represented in the event included the Chinese Academy of Sciences, Tsinghua University, the University of Science and Technology of China, the University of California San Diego, Sber Robotics Center, Alibaba, Amap and vivo. More than 100 teams exceeded the official baseline.
The competition combined online automated evaluation with an offline final using real robots in Vienna. AGIBOT used its EWMBench and Genie Sim Benchmark frameworks to support automated testing, standardized metrics and reproducible evaluation across simulation and physical execution.
T
During the offline final, teams completed tasks using AGIBOT’s G2 humanoid robot in real-world scenarios. The evaluation incorporated robot stability, adaptability in physical environments and long-horizon task reliability as part of the scoring process.
The Reasoning to Action track expanded on the 2025 Manipulation track by covering environment understanding, task planning and physical execution. Teams trained reasoning-and-manipulation models using the AGIBOT World open-source dataset and evaluated them through Genie Sim 3.0. The benchmark covered language understanding, spatial reasoning, atomic skills, disturbance adaptation and zero-shot transfer. PrismBot from vivo won the track, followed by Shanghai RoboParty’s RP-VLA and GreenVLA.
AGIBOT and Dexmal also introduced a real-supermarket benchmark track focused on end-to-end decision-making and whole-body control. The track required models to perform mobile manipulation tasks in a retail-like environment, including autonomous navigation, item picking, transport and placement. Participants’ algorithms controlled real robots through API-based remote control, with tasks subject to physical constraints such as shelf height limits and randomized item placement.
In the World Model track, NeoVerse-ABot, a joint team from the Institute of Automation of the Chinese Academy of Sciences and Amap CV Lab, placed first. PAI@IAII from the Institute of Industrial Artificial Intelligence, Chinese Academy of Sciences, ranked second, and Loop from the University of Science and Technology of China placed third. The track included non-ideal physical interactions such as dropped objects and grasping failures.
AGIBOT also made available a toolchain covering real-world data, simulation evaluation and real-robot testing. The toolchain includes the AGIBOT World open-source dataset, Genie Sim 3.0 and the AGIBOT G2 robot platform.
The company said resources developed through the competition will be integrated into its benchmark development and open-source ecosystem. AGIBOT plans to launch an online simulation leaderboard, add test tasks and benchmarks, and continue developing quantitative evaluation tools for embodied AI systems.
