This competition is built upon the Embodied City platform, led by the Center for Urban Science and Computing, Department of Electronic Engineering, Tsinghua University. The platform provides a high-fidelity simulation and evaluation environment for embodied intelligence in real large-scale urban settings. Through digital twin technology, it reconstructs complex building layouts, diverse street-level environments, and dynamic traffic flows, offering a safe, efficient, and scalable testbed for algorithm validation. It supports frontier research in city-scale spatial reasoning, multimodal video understanding, and agent navigation and decision-making.

This competition adopts a leaderboard-based evaluation mechanism and consists of two tracks:

• Track 1: Urban Video Question Answering
• Track 2: Real-time Interactive and Goal-oriented Navigation

Final rankings will be determined by the organizing committee through unified evaluation on hidden test sets. Participating teams may update their submissions before the deadline to improve their scores. The hidden test sets follow the same data organization, field definitions, and evaluation protocols as the corresponding public benchmarks, allowing participants to develop and adapt their methods using public datasets.

Participating teams may submit either a callable model API or model weights with the corresponding inference code. The organizing committee will conduct unified inference and evaluation. Submissions via callable API will be prioritized in the evaluation queue and receive faster feedback.

Track 1: Urban Video Question Answering

This track is based on UrbanVideo-Bench and focuses on evaluating the spatial perception, temporal reasoning, and question-answering capabilities of multimodal models at the urban scale. Participating teams are expected to validate their algorithms on the urban embodied dataset provided by the organizing committee. Given a city-scale video clip and a question, the model should output an accurate textual answer.

The task simulates continuous first-person observations of an aerial urban agent, such as a drone, during flight. Models need to analyze long-sequence video streams, understand the agent’s trajectory in 3D space, capture environmental changes and event logic, and answer specific spatial or behavioral questions by selecting the correct option.

The dataset contains approximately 5,000 high-quality multiple-choice questions, covering 16 fine-grained task categories, including spatial reasoning, object counting, attribute recognition, and event localization. The data realistically captures dynamic traffic, occlusion relationships, and large-scale spatial displacement in urban environments.

The track uses standard accuracy as the evaluation metric, directly comparing the model’s predicted option with the ground-truth label.

The public leaderboard currently includes baseline results reported in the corresponding paper, which participants may use as references for task difficulty and public model performance. To facilitate unified evaluation by the organizing committee, participating teams are encouraged to provide interfaces that are compatible with the official evaluation workflow.

For the Video QA track, participants may refer to run.py and eval.py in the UrbanVideo-Bench code repository to organize model outputs in the required format. Model outputs should include explicit answer options to support standardized parsing and score calculation.

Track 2: Real-time Interactive and Goal-oriented Navigation

This track is based on EmbodiedNav-Bench and evaluates the closed-loop control and decision-making capabilities of embodied agents in dynamic and complex real-world urban digital twin environments. Participating teams are expected to develop agents that can process multimodal perception data in real time and generate continuous action commands, bridging the gap from “cognition” to “action.”

Participating teams will complete a Vision-Language Navigation task. Given a natural-language description of a target location and real-time environmental observations, the agent should autonomously generate a sequence of valid control commands and successfully reach the designated location.

Participants are required to control a virtual agent deployed in the Embodied City simulation platform, such as an aerial drone. The system provides a natural-language instruction describing the target location. The agent receives first-person RGB images and/or depth maps in real time, performs online decision-making through multimodal models, and outputs executable control signals, including movement, elevation changes, and turning actions, to reach the target area.

The dataset provides standardized samples containing starting-point coordinates, navigation instructions, and corresponding expert reference trajectories. It supports imitation learning, planning-based reasoning, and policy optimization.

The track will be evaluated using the following metrics:

• Success Rate (SR)
• Success weighted by Path Length (SPL)
• Distance to Goal (DTG)

The public leaderboard currently displays baseline results reported in the corresponding paper, which participants may use as references for task difficulty and public model performance. To facilitate unified evaluation by the organizing committee, participating teams are encouraged to provide interfaces that are compatible with the official evaluation workflow.

For the navigation track, participants may refer to embodied_vln.py in the EmbodiedNav-Bench code repository and follow the standard navigation workflow to return executable action commands for unified evaluation and reproducible results.

Timeline and Submission

Submission Deadline: July 29, 2026

Competition 1 adopts a leaderboard-based ranking mechanism. Final rankings and awards will be determined according to the organizing committee’s unified evaluation results on the hidden test sets. Participating teams may update their submissions before the deadline to improve their performance and do not need to resubmit the registration form repeatedly.

Participating teams should submit the following information through the leaderboard platform:

• Team information
• Selected track
• Model name and version
• Submission type: API or model weights with inference code
• A brief technical description

For API-based submissions, teams should provide:

• API endpoint
• Authentication method
• Calling instructions
• Input/output format
• Example request

For model-weight submissions, teams should provide:

• Model weights
• Inference code
• Dependency and environment instructions
• Executable scripts

Registration and result submission for Competition 1 should be completed through the leaderboard platform: https://huggingface.co/spaces/EmbodiedCity/EmbodiedCity-Leaderboard

This competition does not impose a fixed research question, dataset, or evaluation task. Participants may focus on important urban science topics such as urban inequality, human mobility, climate adaptation, public health, spatial equity, infrastructure allocation, and other related issues. They may use open data, remote sensing imagery, street-view images, POIs, transportation data, meteorological data, socioeconomic data, and other multi-source urban datasets. With AI as a research partner, participants are expected to uncover urban patterns that have not been previously documented.

To support efficient AI-assisted research, the organizing committee provides a set of tool resources. At the skill-module level, participants are encouraged to use the series of Skills provided in the AI Urban Scientist repository, which cover the full workflow of urban science research, including research question generation, data search and discovery, analytical strategy recommendation, data analysis, and paper/report writing.

At the platform level, the organizing committee will also provide three types of platform-level tools:

• An Ideation Platform for research topic exploration and idea generation
• A Data Discovery and Download Platform for retrieving and downloading multi-source urban data
• An integrated research platform for supporting the full research workflow from concept development to final output

Participants may also use other AI tools according to their own research needs.

The evaluation of this competition will focus on both the scientific value of the research findings and the human-AI collaborative research process. Key evaluation criteria include:

• Importance and novelty of the research question
• Creativity of data sources and data integration
• Rigor and reproducibility of the analytical process
• Effectiveness of AI collaboration
• Contribution of the findings to urban science

We look forward to seeing unexpected combinations of data, counter-intuitive urban patterns, and discoveries that expand the boundaries of urban science.

Timeline and Submission

Preliminary Review Submission Deadline: July 22, 2026

Participating teams should complete their research using AI tools of their choice and submit the following materials:

1. Research Report
   The report should include the research question, data sources, methodology, main findings, and significance for urban science.
2. Data Description and Reproducible Code
   Participants should provide a clear description of the data used and submit reproducible code where applicable.
3. AI Collaboration Record
   Participants should document the human-AI collaboration process. Acceptable forms include conversation logs, code repository records, or a research log.

The organizing committee will invite experts to review the preliminary submissions, provide feedback to participating teams, and select outstanding teams for the final presentation.

Preliminary Review Results and Feedback: July 29, 2026

Final Presentation: August 8–9, 2026
Venue: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Shortlisted teams will be invited to participate in the Fourth International Conference on Urban Science and Intelligence. During the conference, they will present their research outcomes on site, including the research question, data sources, AI collaboration process, methodology, and main findings. The presentations will be reviewed and discussed by an international panel of expert judges.

Registration and Submission

Registration Link: https://www.wjx.top/vm/OStaaOV.aspx#

All submission materials should be packaged and uploaded to the Urban Cup 2026 cloud folder:

https://cloud.tsinghua.edu.cn/u/d/a0641235ca584f89ad73/

Please name the report file in the following format:

Competition 2_Team Name_Project Title

Example: Competition 2_Test Team_Urban Research Report

Focusing on key challenges in real-world urban operations, this competition encourages participants to use large-scale agent-based simulation, urban data analysis, and human-AI collaborative research methods to develop explainable, verifiable, and reproducible solutions to urban problems.

This competition does not impose a fixed research question, dataset, or evaluation task. Participating teams may use the agent modeling capabilities, urban environment interfaces, and experimental evaluation tools provided by AgentSociety² to design their own research tasks. Possible topics include, but are not limited to, human mobility, urban transportation, public services, disaster response, community governance, sustainable development, and the evolution of social behaviors.

Participants are expected to construct multi-agent urban experimental scenarios and propose analytical, predictive, or intervention-oriented solutions with real-world relevance.

This competition values not only model performance, but also problem formulation, mechanism interpretation, and policy significance. Participating teams are encouraged to investigate core questions such as how urban systems operate, how different groups interact, and what cascading effects may result from specific policies or interventions.

Interdisciplinary teams are strongly encouraged. Cross-university and cross-institution teams are welcome. The organizing committee will set up a dedicated Urban Cup module on the AgentSociety² platform and provide each participating team with a RMB 100 computing voucher, together with platform access and necessary technical support.

The evaluation of this competition will focus on the scientific and practical value of the simulation-based research, including:

• Importance and clarity of the research question
• Appropriateness of the simulation design
• Use of AgentSociety² capabilities
• Mechanism explanation and interpretability
• Rigor, reproducibility, and validity of the analysis
• Policy relevance and real-world implications
• Quality of presentation and communication

Timeline and Submission

Preliminary Review Submission Deadline: July 22, 2026

Participating teams should select their own research topic, conduct the study using the AgentSociety² platform, and submit the following materials:

1. Research Report
   The report should include the research question, data sources, methodology, simulation design, main findings, and significance for urban science.
2. AgentSociety² Workspace
   Teams should provide the corresponding AgentSociety² workspace, project link, or other materials that allow the organizing committee to review the simulation design, experimental setup, and results.

The organizing committee will invite experts to review the preliminary submissions, provide feedback to participating teams, and select outstanding teams for the final presentation.

Preliminary Review Results and Feedback: July 29, 2026

Final Presentation: August 8–9, 2026
Venue: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Shortlisted teams will be invited to participate in the Fourth International Conference on Urban Science and Intelligence. During the conference, they will present their work on site, including the research question, data sources, simulation design, AgentSociety² workflow, and main findings. The presentations will be reviewed and discussed by an expert judging panel.

Registration and Submission

Registration Link: https://www.wjx.top/vm/OStaaOV.aspx#

All submission materials should be packaged and uploaded to the Urban Cup 2026 cloud folder:

https://cloud.tsinghua.edu.cn/u/d/a0641235ca584f89ad73/

Please name the report file in the following format:

Competition 3_Team Name_Project Title

Example: Competition 3_Test Team_Urban Social Simulation Report