Can AI handle the real world? AGIBOT's latest challenge finds out.
AGIBOT used this year's AGIBOT World Challenge 2026 to push embodied AI evaluation beyond virtual environments and into real-world testing, bringing together 526 research and enterprise teams from 27 countries during ICRA 2026.
The Shanghai-based robotics company designed the competition around a growing concern in the industry: whether AI systems that perform well in simulations can deliver the same results when deployed on physical robots. Instead of relying solely on benchmark scores generated in virtual settings, the challenge incorporated real robots, real tasks, and standardized evaluation methods.
Finalists were required to complete tasks using the AGIBOT G2 humanoid robot during an in-person final held in Vienna. The approach placed greater emphasis on factors such as stability, adaptability, and long-term task execution, qualities that are critical for real-world deployment but often difficult to measure through simulation alone.
Two tracks tested core AI capabilities
The competition was divided into two categories that targeted different aspects of embodied intelligence.
The Reasoning to Action (R2A) track examined how robots interpret instructions, understand their surroundings, develop plans, and carry out tasks in physical environments. The category represents an expansion of AGIBOT's previous manipulation-focused evaluation, broadening the assessment from simple task execution to a complete reasoning-and-action pipeline.
The World Model (WM) track focused on prediction and environmental understanding. Teams were challenged to build systems capable of forecasting how physical environments would change in response to robot actions and sensor data.
According to AGIBOT, the dual-track structure reflects the industry's progression from teaching robots how to perform tasks toward enabling them to understand, predict, and make decisions in dynamic environments.
Participation came from a mix of universities, research institutions, startups, and technology companies. More than 100 teams exceeded the competition's baseline performance requirements.
vivo team takes top honors in reasoning category
Participants in the R2A track trained models using the AGIBOT WORLD open-source dataset and tested them through Genie Sim 3.0. The benchmark measured several capabilities, including language comprehension, spatial reasoning, disturbance handling, atomic manipulation skills, and zero-shot transfer performance.
PrismBot, a team from vivo, secured first place in the final standings. Shanghai RoboParty's RP-VLA finished second, while GreenVLA claimed third place.
Real supermarket becomes a testing ground
One of the event's more practical additions was a supermarket benchmark developed jointly by AGIBOT and Dexmal.
The benchmark placed robots inside a retail environment where they had to navigate aisles, locate products, pick items, transport them, and place them in designated locations. Teams had to operate under real-world constraints such as shelf-height limitations and randomized product placement.
Rather than running purely in simulation, participants remotely controlled physical robots via APIs, enabling organizers to evaluate how algorithms performed under realistic deployment conditions.
The benchmark also introduced challenges such as dropped objects and failed grasps, creating conditions that more closely resemble real-world robotic operation.
Chinese Academy of Sciences team leads world model rankings
The World Model competition produced a different set of winners. NeoVerse-ABot, a joint team from the Institute of Automation of the Chinese Academy of Sciences and Amap CV Lab, captured first place.
The PAI@IAII team from the Institute of Industrial Artificial Intelligence at the Chinese Academy of Sciences earned second place, while the Loop team from the University of Science and Technology of China finished third.
The track focused on evaluating how effectively AI systems could model and anticipate physical-world interactions, particularly in the presence of unexpected events.
AGIBOT opens its development stack
Beyond the competition, AGIBOT also made available a broader development framework intended to support embodied AI research.
The stack includes the AGIBOT WORLD open-source dataset, Genie Sim 3.0 simulation platform, and the AGIBOT G2 humanoid robot platform. Together, these tools are designed to help developers train, evaluate, and validate robotic AI systems from simulation through physical deployment.
The company said resources developed through the challenge will feed into its ongoing benchmark and open-source initiatives. Future plans include launching an online simulation leaderboard, adding new testing tasks, and expanding benchmark coverage to provide more comprehensive measurements of embodied AI performance.
Also read: Unitree’s IPO review could add another public-market test for China’s humanoid robotics sector.