The event centers on real-robot testing with standardized benchmarks, moving from simulations to closed-loop evaluation on real robots and tasks using the AGIBOT G2 humanoid robot in offline final assessments.

There is a push for practical deployability, robustness, and long-horizon task reliability to bridge the gap between simulation scores and real-world performance.

EWMBench and Genie Sim Benchmark were showcased to standardize evaluation, enabling reproducible results across simulation and physical testing and improving robustness, deployability, and generalization of embodied AI models.

AGIBOT WORLD CHALLENGE 2026, held with ICRA 2026 in Vienna, drew 526 teams from 27 countries to compete in two embodied AI tracks: Reasoning to Action and World Model.

In the Reasoning to Action track, teams trained on the AGIBOT WORLD open-source dataset and were evaluated with Genie Sim 3.0 on language understanding, spatial reasoning, atomic skills, disturbance adaptation, and zero-shot transfer; PrismBot (vivo) won, followed by RP-VLA (Shanghai RoboParty) and GreenVLA.

Future plans include an online simulation leaderboard, more test tasks and benchmarks, expanded quantitative evaluation, and ongoing development of benchmarks and the open-source ecosystem with global partners.

Beyond competition, AGIBOT aims to launch an online simulation leaderboard, add more test tasks and benchmarks, and continue developing benchmarks and the open-source ecosystem to advance embodied intelligence toward real-world deployment.

AGIBOT unveiled a full-stack toolchain—AGIBOT WORLD dataset, Genie Sim 3.0, EWMBench, Genie Sim Benchmark, and the AGIBOT G2 platform—to enable cross-stage validation from training to deployment with standardized metrics across simulation and real-robot testing.

In the World Model track, top teams emphasized robustness under non-ideal interactions, with NeoVerse-ABot (CAS and Amap CV Lab) first, PAI@IAII second, and Loop (USTC) third.

The R2A track expanded from manipulation to full environment understanding, planning, and physical execution, with PrismBot leading, then RP-VLA and GreenVLA among the top contenders.

A real-supermarket benchmark track was introduced to test end-to-end decision-making and whole-body control in a realistic retail environment, with autonomous navigation, item picking, transport, and placement under real physical constraints.