AGIBOT World Challenge at ICRA 2026 Now Opens "Reasoning to Action" and "World Model" Tracks

SHANGHAI, March 3, 2026 /PRNewswire/ -- AGIBOT, a global leader in embodied intelligence and robotics, today officially announced the launch of the AGIBOT World Challenge at ICRA 2026, opening registration for its two flagship tracks: the Reasoning to Action Track and the World Model Track. Following the success of its inaugural edition, this year's competition returns to the IEEE International Conference on Robotics and Automation (ICRA) with a broader vision and a record-breaking $530,000 USD prize pool.

Cutting-Edge Tracks, Ultimate Challenges

1. Reasoning to Action Track

This track is designed to evaluate a model's reasoning and action execution capabilities, encompassing both online simulation and offline real-robot stages. Participants will train models capable of solving a series of complex tasks based on the AGIBOT WORLD open-source dataset. The track focuses on bridging the Sim2Real Gap, achieving robust generalization from open-vocabulary understanding to real-world physical interaction. Tasks are designed in conjunction with real deployment scenarios and the Genie Sim 3.0 simulation platform, spanning common environments such as logistics, industrial, supermarket, food service, and home settings — with multiple challenges of varying difficulty levels.

Featured Task Scenarios:

• Continuous Sorting & Barcode Scanning
• Logistics Sorting
• Workpiece Inversion
• Shelf Stocking & Straightening Tilted Items
• Popcorn Scooping
• Door Opening
• Desk Clearing
• Dual-Arm Pot Handling

Dataset

We have carefully curated competition tasks and collected a dedicated set of high-quality data, with each task containing hundreds of complete manipulation trajectories. The dataset is fully open-sourced on Hugging Face and ModelScope for easy download:
https://huggingface.co/datasets/agibot-world/AgiBotWorldChallenge-2026
https://modelscope.cn/datasets/AgiBotWorld/AgiBotWorldChallenge-2026

Baseline Model

The competition will provide ACoT-VLA as the baseline model to help you quickly get up to speed with the training, testing, and submission workflow. You can compare the performance of different baseline models and select the most suitable one for development. We look forward to seeing your innovations built upon these models!
https://github.com/AgibotTech/ACoT-VLA 

Evaluation Tool

For model evaluation, we provide Genie Sim 3.0 — the industry's first open-source simulation platform driven by large language models. This comprehensive evaluation platform integrates 3D reconstruction and visual generation to create a photorealistic, digital-twin-grade environment. Scenes, assets, and tasks within Genie Sim 3.0 are consistent with the test server, enabling participants to perform closed-loop evaluation locally. Genie Sim 3.0 utilizes AGIBOT's proprietary embodied brain Genie Reasoner to achieve fully automated VLM-based evaluation. Before submitting to the competition server, you can first validate your model and environment dependencies in the local simulation environment.
https://agibot-world.com/genie-sim

Competition Server

Starting now, participants can log in to the competition homepage to access event resources and submit their trained models — compete directly with the world's top developers and keep climbing up the ranking list!competing head-to-head with top developers worldwide and climbing the leaderboard!
https://agibot-world.com/challenge2026/reasoning2action/quick-start

2. World Model Track

The World Model Track focuses on the core capability of embodied world models — accurately modeling physical environment dynamics based on robot actions. Participants are required to train video generation models on the AGIBOT WORLD dataset, generating interaction videos of robots operating across 10 real-world task scenarios, based on provided real robot observations and action signals. Tasks span multiple real-world environments including furniture, retail, and industrial settings, providing a comprehensive evaluation of model generation quality. This track features an online-only competition stage.

Dataset

We have constructed a World Model training set of 10 diverse tasks based on the AGIBOT WORLD public dataset, covering over 30,000 real trajectories. The data encompasses a wide variety of robot-environment interaction types, such as grasping, placing, pushing, and pulling. The test set includes both expert trajectories and imperfect action trajectories (e.g., missed grasps, collisions), designed to comprehensively evaluate the generalization capability of generative models.
https://huggingface.co/datasets/agibot-world/AgiBotWorldChallenge-2026/tree/main/WorldMode 

Baseline Model

The competition adopts EVAC — AGIBOT's proprietary open-source model — as the baseline. EVAC is the first embodied world model driven by robot actions, pre-trained on the full AGIBOT WORLD dataset. Participants may develop on top of EVAC or use other models to compete. We encourage participants to innovate boldly and showcase their technical capabilities!
https://github.com/AgibotTech/AgiBotWorldChallengeICRA2026-WorldModelBaseline 

Evaluation Tool

The competition evaluates generated results using EWMBench, an embodied world model evaluation benchmark, assessing across multiple dimensions including image quality, scene consistency, and trajectory adherence — accurately reflecting generative model performance and providing reliable feedback.
https://github.com/AgibotTech/EWMBench

Evaluation Server

Starting now, participants can log in to the evaluation server hosted on HuggingFace, submit results according to the competition guidelines, and compete with developers worldwide — with the leaderboard updated in real time!
https://huggingface.co/spaces/agibot-world/ICRA26WM

Key Milestones — Mark Your Calendar

The AGIBOT World Challenge @ ICRA 2026 follows a rigorous selection process, culminating in the live finals this summer:

• February 12, 2026: Global Registration officially opens
• February 28, 2026: Competition servers for both tracks go live
• April 20, 2026: Competition server closes for final submissions
• April 30, 2026: Announcement of online stage results and finalists
• June 1, 2026: Offline real-robot finals begin at ICRA 2026

The stage is set for the world's most elite teams to demonstrate their prowess. As the AGIBOT World Challenge at ICRA 2026 unfolds, it will serve as a definitive arena where technical ingenuity meets physical capability, and where the strongest innovators will rise to define the next era of robotics.

Join the Community

AGIBOT invites developers and researchers worldwide to join the official Lark (Feishu) community. Connect with global peers, access real-time technical support, and stay updated on the latest challenge milestones through the following link: https://applink.feishu.cn/client/chat/chatter/add_by_link?link_token=b03mea0a-1212-428a-8b78-f77cf6f591e3

For more information, please visit AGIBOT at AGIBOT.com and follow AGIBOT on:
https://www.facebook.com/AGIBOT.zhiyuan
https://x.com/AGIBOT_zhiyuan
https://www.instagram.com/AGIBOT
https://www.youtube.com/@AGIBOT-robot 

About AGIBOT

AGIBOT is dedicated to driving innovation through the integration of AI and robotics, creating world-leading general-purpose embodied robot products and an application ecosystem. Built on the foundation of the robotic body and powered by the fusion of interaction, manipulation, and locomotion intelligence — "1 Robotic Body + 3 Intelligence" — AGIBOT is a leading robotics company in the industry to deliver a complete product portfolio and deploy across all major application scenarios.

SOURCE Agibot