Starting from 2,999 Submissions: Observations on AI Innovation across GOAI's Four Tracks

The preliminary-round submissions of the Global Open-source AI Challenge (GOAI) cover the fields of enterprise-grade agents, industry applications, AI scientific research, and embodied intelligence. Although the four tracks have different directions, they share a common thread: participating teams no longer focus only on whether a model "can do it," but begin to answer whether the system can truly run, complete tasks, handle failures, and prove results with evidence.

From the 2,999 preliminary-round submissions, we can not only see a number of high-quality AI innovation projects, but also observe ongoing changes in agent technology, product forms, and the open-source ecosystem.

What innovation directions do the preliminary-round submissions present?

Track 1 "Agent Infra": Multi-Agent Systems Move Toward Enterprise Systems

The Agent Infra track covers software R&D, Agent infrastructure and SRE governance, financial risk control, industrial manufacturing, scientific research, and enterprise operations. From the preliminary-round submissions, we can see that participating teams are moving away from generic Q&A and simple Demos toward solving complex tasks in real enterprise scenarios. Submissions mainly adopt technical paths such as hierarchical multi-Agent collaboration, dynamic orchestration, independent verification, human gates, Skill and tool integration, exception recovery, and version governance.

These submissions attempt to solve problems in complex enterprise tasks such as fragmented information, difficulty in directly trusting AI results, lack of control over high-risk actions, and difficulty in reusing organizational experience. Compared with single agents, teams are more concerned with how multiple Agents divide work, how tools are invoked, how processes are audited, and how results are verified.

Track 2 "Boundless Agents": AI Enters Real Workflows

The Boundless Agents track presents a rich and specific application landscape, covering industry, finance, automotive, education, smart glasses, healthcare and elderly care, enterprise collaboration, and personal services.

The preliminary-round submissions can be broadly categorized into industry-vertical Agents, enterprise process automation, personal and accessibility assistants, Agent infrastructure and governance, and developer and open-source tools. For example, industrial projects enter equipment diagnosis, quality management, and work-order collaboration; financial projects focus on risk control and compliance, investment research analysis, and trading assistance; automotive projects revolve around trip planning, vehicle maintenance, and safety emergencies; education projects cover teaching content generation, vertical learning, and research assistance; smart glasses projects attempt to use first-person perspective for recognition, reminders, and communication assistance, or connect continuous perception with Agent tasks. Different submissions collectively point to one goal: moving AI from "answering questions" to "completing tasks."

Track 3 "AI for Research": AI Participates in Scientific Research

The AI for Research track includes algorithm sub-tracks and open exploration sub-tracks, with submissions covering AI virtual cells, materials science, life sciences, quantum computing, climate and environment, and social sciences and humanities, demonstrating distinct research attributes and interdisciplinary characteristics.

Algorithm sub-track virtual cell projects attempt to use statistical baselines, low-rank decoding, residual decomposition, and generative models to address measurement bias and generalization to new entities. Materials research projects revolve around literature retrieval, knowledge graphs, Research Gap identification, and candidate material generation, while attempting to introduce physical constraints and database verification.

Open Exploration submissions demonstrate the breadth of AI integration with multiple disciplines, as participants begin to involve AI in problem definition, scientific reasoning, hypothesis generation, and result verification.

Track 4 "Embodied Future": Agents Enter the Physical World

The Embodied Future track includes Patrol and Dual-arm sub-tracks. Patrol submissions focus on sensor-control fusion, environmental mapping, autonomous navigation, and task execution. Dual-arm submissions mainly follow paths such as vision-language-action foundation model adaptation, multi-view and 3D perception, and scene generalization. Embodied Future projects not only need robots to "understand the environment," but also to connect perception, planning, control, and action into a continuous closed loop.

What Do Judges See in the Submissions?

Agent Infra: Agent Count ≠ System Capability

The most commendable change in this track's preliminary-round submissions is that participating teams have begun shifting from generic Q&A and simple Demos to complex real-enterprise tasks, and started focusing on responsibility separation, independent verification, human approval, rollback auditing, and Skill reuse. This shows that the understanding of enterprise-grade Agent systems is shifting from "whether the model can complete the task" to "how the system reliably completes tasks within a real chain of responsibility." The key evaluation criterion is not how many Agents the system has, but whether these Agents undertake real and necessary collaboration.

Boundless Agents: Leading Submissions Have Started to "Get Things Done"

The most commendable change in the preliminary round is that many submissions are no longer satisfied with "AI being able to give suggestions," but begin to think about how to act after the suggestions. Outstanding submissions do not seek to cover all scenarios, but choose a real, clear pain point and connect recognition, decision-making, execution, and feedback into a complete closed loop, enabling users to actually experience it, and clearly distinguish which tasks can be handed to AI and which actions must be confirmed by humans. Some automotive and financial submissions have also begun to clarify AI's capability boundaries, setting human gates when involving funds, safety, and external operations.

AI for Research: Scientific Claims Must Withstand Verification

It is commendable that outstanding Algorithm sub-track submissions have begun to focus on the structure of scientific problems themselves, emphasizing controlled experiments, data sources, measurement conditions, and research reproducibility. Outstanding Open Exploration sub-track submissions can start from real professional problems and let AI actually participate in analysis, reasoning, and verification.

Embodied Future: From Model Demonstration to Problem Orientation

In the Embodied Future track, some teams no longer blindly pursue larger models, but instead improve on issues such as inaccurate grasping, unstable placement, lighting reflections, positioning deviations, and inability to recover after failures. Some teams have completed large-scale simulation verification and provided training and inference tools, demonstrating strong engineering capability and result awareness. However, some projects still need to improve their testing.

What Industry Trends Do the Competition Submissions Reflect?

Agent Infra submissions reflect that enterprise-grade AI is shifting from model capability competition to system capability competition. The key to the future is organizing models, Agents, Skills, tools, data, and human responsibility into reliable systems. Observability, verifiability, recoverability, and auditability will become basic requirements for production-grade agents.

Boundless Agents submissions show that AI products are shifting from "adding an intelligent feature" to "taking on a specific process." Large models handle understanding, planning, and coordination, while databases, rule engines, and specialized tools handle deterministic execution. Future products need more "narrow scenarios, deep services," going deep and stable on one real problem.

AI for Research submissions indicate that the competitive focus of research AI is shifting from model scale and Agent count to data quality, domain constraints, and verification capability. Only by establishing a chain from evidence, hypothesis, and prediction to independent verification can AI truly enter the scientific discovery process.

Embodied Future submissions reflect that robot competition is moving from individual algorithms to system integration of perception, decision-making, control, and hardware. As vision-language-action foundation models gradually become a common starting point, data quality, multi-view perception, key action learning, closed-loop error correction, and real-environment adaptation will become the main sources of differentiation.

The submissions of four tracks also show a common open-source trend: the scope of openness is expanding from models and code to Agents, Skills, interface protocols, evaluation tools, and deployment specifications. Open-source standards are also shifting from "code can be seen" to "projects can be installed, run, and reproduced."

Entering the Semi-Finals, Which Capabilities Deserve Focus?

Agent Infra will focus on verifying whether Agents, Skills, and tools can form a complete business chain. Beyond normal workflows, projects must also demonstrate whether the system can safely stop, recover, or replan when facing unavailable tools, data conflicts, and human rejections.

Boundless Agents needs to prove three outcomes: real problem, real capability, and real closed loop. Judges will not only look at one successful Demo, but also focus on whether the system can run stably when facing unfamiliar inputs and tool failures, and whether business value can be verified through data, controlled experiments, or user feedback.

AI for Research will focus on observing whether submissions can truly run, whether core capabilities can be independently verified, and whether scientific conclusions have clear and credible evidence boundaries. Algorithm sub-track needs to prove that the model has truly learned entity-specific changes; Open Exploration sub-tracks need to explain whether AI truly participates in core problem solving and whether interdisciplinary integration has substantive depth.

The focus of Embodied Future track's final round will shift from simulation to real robots. Dual-arm robots need to prove collaborative operation, safety control, and failure recovery capabilities; patrol robots need to verify the continuous closed loop of perception, mapping, navigation, obstacle avoidance, and task execution. Beyond success rate, task duration, number of human interventions, number of exceptions, recovery effectiveness, and multi-round performance fluctuation will also become important metrics.

AI Innovation Is Shifting from Capability Demonstration to Real Verification

GOAI's four tracks explore four directions of AI innovation respectively: Agent Infra builds reliable agent infrastructure, Boundless Agents pushes AI into real business processes, AI for Research lets AI participate in scientific research, and Embodied Future brings agents into the physical world.

The four paths ultimately converge on the same set of standards: whether the project solves real problems, whether core capabilities can be verified, whether the system can run stably, and whether the results can be externally reproduced and co-built.

Moving from "making an AI project" to "proving it is real, reliable, and has lasting value" will become the true dividing line in the GOAI semi-final stage.

Scan to Add GOAI Assistant

Your Exclusive Competition Guide Is Now Online