A Global Open-Source AI Contest Enters Review

The World Artificial Intelligence Open Source Competition, known as GOAI, has started its preliminary review after submissions closed for its first three tracks at 23:59 Beijing time on August 16.

The review covers Track 1, Agent Infra; Track 2, Boundless Agents; and Track 3, AI for Research. A professional judging panel will evaluate 2,899 submitted projects from around the world. Track 4, Embodied Future, remains open for registration and submission until 23:59 Beijing time on August 20.

According to the organizers, GOAI has attracted participants from 91 countries and regions across six continents. More than 12,279 teams and over 14,000 participants have registered, including 1,549 international teams and 1,749 international participants. The field includes AI builders, universities, research institutions, enterprise engineering groups and startup teams.

In this context, “AI builders” refers to developers and teams that create AI models, tools, infrastructure and applications. For an open-source competition, the breadth of participation matters because it brings different technical approaches and regional use cases into a common review process.

Four Tracks Reflect the Current AI Agenda

GOAI’s four tracks map closely to several major directions in AI development: agent infrastructure, industry applications, AI for science, and embodied intelligence.

The main participation figures are:

• Agent Infra: 847 preliminary submissions, with 7,802 registered teams and 8,346 participants. The track focuses on enterprise-grade multi-agent infrastructure and collaboration systems.
• Boundless Agents: 1,280 submissions, with 2,460 teams and 3,160 participants. Submitted projects cover areas such as education, finance, industrial manufacturing and smart glasses.
• AI for Research: 607 submissions, with 1,495 teams and 1,877 participants. The track encourages runnable algorithmic solutions or autonomous exploration environments for real scientific questions.
• Embodied Future: 165 submissions so far, with 522 teams and 670 participants. This track focuses on open-source embodied intelligence and interaction with the physical world, and is still accepting work.

An AI agent is software that can interpret a task, use tools and carry out steps toward a goal. A multi-agent system coordinates several such agents. Embodied intelligence refers to AI systems that interact with physical environments, such as robotic manipulation or patrol tasks.

How the Preliminary Review Works

The preliminary round uses an online independent review mechanism with experts from industry, academia and research. Projects will be assessed on technical innovation, real-world or research value, demo completeness, open-source contribution, ecosystem value, engineering maturity and the team’s long-term potential.

The criteria differ by track. Agent Infra emphasizes scenario value, industry replicability, multi-agent collaboration, autonomous closed-loop capability, skill engineering and operational validation. Boundless Agents focuses on industry value, agent capability, product experience, engineering reproducibility, safety and compliance. AI for Research has separate criteria for its algorithm and open exploration tasks, weighing factors such as technical performance, scientific significance, problem definition and verifiability. Embodied Future relies mainly on online simulation evaluation, including task quality and system performance for dual-arm tasks, and route completion time for patrol tasks with different coefficients for autonomous navigation and remote control.

Before expert scoring, the organizing committee will check whether submissions match the task requirements, whether materials are complete and whether the entries comply with the rules. Qualified projects will then be anonymized for expert review. Each project is generally scored independently by two to three judges, and the final score is the arithmetic average of valid scores. Additional review may be triggered if there is a conflict of interest, a large scoring gap or a major disagreement about eligibility.

Advancement and Upcoming Milestones

For each track, the top 300 projects will receive up to 200 yuan in resource subsidies and a commemorative competition T-shirt. In terms of advancement, the top 30 teams from Agent Infra and the top 30 from Boundless Agents will enter the next round. AI for Research will select 20 teams from each of its two tasks, for a total of 40 teams. Embodied Future will select 10 teams from each of its two tasks, for a total of 20 teams entering final preparation.

For Tracks 1, 2 and 3, preliminary review runs from August 17 to August 24, with finalist lists for the next round announced on August 24. Teams will refine and submit their next-round work from August 25 to September 3. The second review runs from September 4 to September 10, and the final-round list will be announced on September 10.

For Track 4, submissions close on August 20. Preliminary review runs from August 21 to August 23, and the offline debugging list will be announced on August 24. All tracks will move toward offline final presentations and demonstrations on September 22, followed by the champion project Showtime, grand prize selection and awards ceremony on September 23.

The broader significance of GOAI lies not only in the ranking results, but in whether open-source AI projects can move from impressive demos to reproducible engineering and validated scenarios. AI agents, AI for science and embodied intelligence are all in a period of rapid experimentation. By putting innovation, open-source contribution and real task performance into one review framework, the competition may help identify teams that are closer to practical deployment. The next rounds will show which projects can keep running reliably beyond the first submission.