AGIBOT World Challenge 2026 Tests Humanoid Robots in Real-World Settings
The AGIBOT World Challenge 2026 tested 526 teams from 27 countries using real-world tasks for humanoid robots, shifting focus from simulations to physical deployment in Vienna. Winners included vivo’s PrismBot in the Reasoning to Action track and a team from the Chinese Academy of Sciences in the World Model category, with a supermarket benchmark introducing realistic constraints like dropped objects and randomized product placement.
AGIBOT’s 2026 World Challenge moved humanoid robot testing from virtual environments to real-world scenarios, involving 526 teams from 27 countries during ICRA 2026. The competition, hosted by Shanghai-based AGIBOT, aimed to address industry concerns about AI performance gaps between simulations and physical deployment. Finalists used the AGIBOT G2 humanoid robot to complete tasks in Vienna, emphasizing stability, adaptability, and long-term execution—qualities difficult to measure in simulations. The challenge featured two tracks: Reasoning to Action (R2A) tested robots’ ability to interpret instructions, plan actions, and execute tasks in physical environments, expanding beyond simple manipulation. The World Model (WM) track focused on predicting environmental changes based on robot actions and sensor data, reflecting the industry’s shift toward dynamic decision-making. Over 100 teams surpassed baseline requirements, with participants including universities, startups, and tech companies. In the R2A track, vivo’s PrismBot secured first place, followed by Shanghai RoboParty’s RP-VLA and GreenVLA. Teams trained models using AGIBOT WORLD’s open-source dataset and tested them via Genie Sim 3.0, evaluating language comprehension, spatial reasoning, and zero-shot transfer performance. A supermarket benchmark, developed with Dexmal, required robots to navigate aisles, locate products, and handle real-world constraints like shelf heights and randomized placements. Teams remotely controlled physical robots via APIs, exposing algorithms to challenges such as dropped objects and failed grasps. The benchmark aimed to simulate realistic deployment conditions beyond simulated environments. In the WM track, NeoVerse-ABot—a joint team from the Chinese Academy of Sciences’ Institute of Automation and Amap CV Lab—won first place. The dual-track structure highlighted the industry’s progression from task execution to environmental understanding and autonomous decision-making in dynamic settings.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.