The robotics industry is fundamentally changing how it measures progress, shifting from theoretical benchmarks in simulation to hands-on testing with real robots performing actual tasks. AGIBOT's World Challenge 2026, held alongside the International Conference on Robotics and Automation (ICRA) in Vienna, brought together 526 research and enterprise teams from 27 countries to compete on embodied artificial intelligence (AI) tasks using physical robots rather than computer simulations alone.

This shift matters because it addresses a critical gap in robotics development: a model might perform perfectly in a simulated environment but fail when a robot encounters the unpredictable messiness of the real world. By requiring teams to validate their AI models on actual hardware, AGIBOT is pushing the industry toward solutions that can genuinely work in factories, warehouses, and retail environments.

What Are the Two Main Competition Tracks Testing?

The challenge featured two distinct evaluation tracks designed to test different aspects of embodied AI capabilities. The "Reasoning to Action" (R2A) track evaluated how robots understand tasks, plan their movements, and execute actions in physical spaces. This track was upgraded from the previous year's manipulation-only focus to include the full process of understanding an environment, planning a strategy, and physically carrying out the work.

The "World Model" (WM) track focused on a different challenge: how AI systems predict what will happen in the physical world based on robot actions and sensor inputs. Teams trained their models using AGIBOT's open-source dataset and tested them through a benchmark called Genie Sim 3.0, which evaluated language understanding, spatial reasoning, basic manipulation skills, the ability to adapt to unexpected disruptions, and zero-shot transfer (performing tasks the model was never explicitly trained on).

Beyond the main competition, AGIBOT and a partner company launched a specialized supermarket benchmark track that incorporated real-world complications like objects being dropped or grasping failures. This track required robots to complete the entire mobile manipulation process, from navigating autonomously to picking items, transporting them, and placing them on shelves with height constraints and randomized item locations.

How Is AGIBOT Addressing the Simulation-to-Reality Gap?

One of the biggest challenges in robotics is that models trained in simulation often perform poorly when deployed on real robots. AGIBOT addressed this by releasing a full-stack toolchain that covers the entire development pipeline:

• Real-world data: The AGIBOT WORLD open-source dataset provides training material grounded in actual robot interactions, not purely synthetic scenarios.
• Simulation evaluation: Genie Sim 3.0 and EWMBench provide standardized metrics and automated evaluation, allowing developers to test models consistently before moving to physical hardware.
• Real-robot testing: The AGIBOT G2 humanoid robot platform enables final validation on actual hardware, ensuring models work in practice, not just in theory.

The benchmarks themselves are designed to support standardized metrics, automated evaluation, and comparable results across both simulation and physical testing. This addresses a longstanding problem in robotics: different teams use different evaluation criteria, making it difficult to compare progress across the industry.

What Do the Competition Results Reveal About Current Capabilities?

The competition results show that more than 100 teams surpassed the official baseline, indicating broad progress in embodied AI. In the R2A track, PrismBot from vivo won the championship with a score of 43.47 points, followed by Shanghai RoboParty's RP-VLA with 35.66 points and Russia's GreenVLA with 33.19 points. 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, took first place.

These results represent teams from leading research institutions and companies, including Tsinghua University, the University of Science and Technology of China, the University of California San Diego, Sber Robotics Center, Alibaba, and vivo. The geographic and institutional diversity suggests that embodied AI progress is not concentrated in a single region or company but is advancing globally.

Why Is This Shift From Simulation to Real-World Testing Important?

The move toward real-robot validation reflects a broader maturation of the robotics industry. For years, progress was measured primarily through simulation benchmarks, which are faster and cheaper to run but often don't capture the complexity of physical deployment. By placing robot stability, real-world adaptability, and long-horizon task reliability at the center of the scoring system, AGIBOT is aligning technical evaluation with practical deployment needs.

"The question is whether you can afford to make any serious technology decision in 2026 without understanding what is happening in this country," said Lorenzo, CEO of IDC, during a keynote presentation at IDC Directions Beijing 2026.

Lorenzo, CEO at IDC

This comment reflects the broader context: AGIBOT shipped more humanoid robots in 2025 than any other company worldwide, amid an 800 percent global market growth last year. The company is actively deploying robots across seven major productivity scenarios, including industrial loading and unloading, logistics sorting, retail service stations, security inspection, and commercial cleaning.

What's Next for Embodied AI Evaluation?

AGIBOT plans to integrate the technical and ecosystem resources developed through the competition with its ongoing benchmark development and open-source efforts. The company intends to launch an online simulation leaderboard, introduce more test tasks and diversified benchmarks, and support more comprehensive quantitative evaluation of model capabilities. The goal is to help embodied AI move from individual algorithmic advances toward systems that can be deployed and scaled in real-world settings.

Wang Chuang, Partner at AGIBOT, outlined what the company calls the "X-Y-Z Curve" of embodied AI industry development. According to this framework, 2026 marks the transition from the X Curve (Development and Exploration Phase) to the Y Curve (Deployment and Growth Phase), officially marking the inaugural year of the deployment era. At this stage, robots will begin to truly work autonomously, serving as the starting point for future explosive growth.

The shift from simulation-based benchmarks to real-robot testing represents a fundamental change in how the robotics industry measures progress and validates technology. By requiring teams to prove their models work on actual hardware performing practical tasks, AGIBOT is pushing the industry toward solutions that can genuinely improve productivity in warehouses, factories, and retail environments. This approach also creates a more level playing field for evaluation, allowing researchers and companies worldwide to compare their progress using standardized metrics and reproducible results.