GATSim: Generative-Agent Transport Simulation

GATSim is a novel framework that leverages generative agents 🤖 to simulate urban transportation systems. Unlike traditional rule-based approaches, GATSim creates autonomous agents capable of human-like decision-making in complex mobility scenarios.

This is an ongoing project that is continuously being improved. It is open-sourced under the Apache 2.0 license, and contributions from practitioners are warmly welcome!

How to cite

Liu, Q., Li, C. and Ma, W., 2026. GATSim: Urban Mobility Simulation with Generative Agents. Transportation Research Part C: Emerging Technologies. Vol 186, 105576.

Liu, Q., Li, C. and Ma W., 2026. Generative Agents for Urban Mobility: A Cognitive Framework for Realistic Travel Behavior Simulation. Simulation Modelling Practice and Theory. Vol 147, 103234.

Overview

This prototype implementation enables generative agents 🤖 in urban simulation environment to:

• Perceive surrounding events/family members & friends state/traffic conditions;
• Interact with family members, friends, and other agents
• Remember with long-term/short-term memory;
• Plan daily activities and adapt schedules dynamically;
• Reflect on travel experiences to improve future decisions;
• React to to unexpected events and changing traffic patterns by re-routing, changing departure time, revising activity plan, etc.

This project is inspired by the work of Park et al. (2023) Generative Agents.

The framework of GATSim:

Motivation

Agent-based simulation has long been a cornerstone of complex systems modeling. Traditional frameworks like MATSim rely on predefined behavioral rules and utility functions to simulate human mobility patterns. While effective, these rule-based approaches struggle to capture the nuanced, adaptive nature of human decision-making. GATSim represents a paradigm shift by employing Large Language Model (LLM)-powered agents that can:

• Make context-aware decisions without explicit programming
• Learn from experiences and adapt behaviors naturally
• Generate realistic responses to novel situations
• Exhibit emergent behaviors that mirror human mobility patterns

Applications

GATSim enables new possibilities for:

• Transportation Planning: Test infrastructure changes and policy interventions
• Traffic Management: Evaluate real-time traffic control strategies
• Behavioral Studies: Understand how people adapt to new mobility options
• Scenario Analysis: Simulate responses to disruptions or emerging technologies
• Urban Design: Assess the impact of land-use changes on travel patterns

By bridging the gap between individual intelligence and collective dynamics, GATSim offers a more naturalistic approach to understanding and optimizing urban mobility systems.

Contributors

Qi Liu, Can Li, Wanjing Ma

MAGIC Lab, College of Transportation, Tongji University

✨Key Features

👉 See more technical highlights at Technical Highlights

👉 See research results at Research Results

Agent in this prototype supports the the following functions:

Concurrent LLM call is supported to speed up simulation process.

💡Agent Long-term Activity Planning Process Each agent generates an initial activity plan at the beginning of the day. The overall planning process is illustrated below. The agent first receives contextual inputs, which include (1) long-term reflections from the previous week, (2) daily reflections from the previous day, and (3) perceived broadcast events

Using this context, the agent may initiate conversations with relevant parties to collaboratively plan the day. For instance, a husband agent might discuss school pick-up responsibilities with his spouse, or an agent might coordinate a post-work gym session with a friend.

The final step in the process involves reasoning and planning. The planning phase outputs three elements: (1) an initial activity plan, (2) a set of concept nodes, and (3) an updated long-term reflection.

🤖Agent Reaction to Real-time Events Process One of the key advantages of generative agents is their ability to flexibly and naturally respond to real-time traffic conditions and unexpected incidents.

🧠Agent Reflection Process Reflection is the process through which an agent internalizes its daily experiences, gradually forming preferences, habits, and abstract knowledge.

Long-term reflections summarize key personal experiences over the past week, encompassing work-related tasks, social interactions, and family affairs. These reflections also contain brief anticipations of upcoming events.

📺 Demo Video

Watch Demo Video

Examples of agent flexible reactive behaviors

Name: George Wright | Time: Thu 2025-03-13 07:28
 I experienced a 9-minute wait at Metro_1_link_1 this morning, which was unexpected and slightly delayed my arrival at the Office. While there is currently only light congestion on Ave_2_link_2 and moderate congestion on Ave_1_link_1, the earlier-than-expected delays on Metro_1 suggest potential variability in transit reliability today. This could affect my ability to comfortably return from the Museum visit during lunch if similar conditions occur on my outbound trip back to Office. However, since no major disruptions are reported and my travel mode remains consistent, I will maintain the current plan with minor timing adjustments.
Plan revisions: George Wright update the departure time of next activity at Museum to 12:45

Name Michael Chen | Time: Thu 2025-03-13 14:05
I am currently at Supermarket, having arrived at 12:40 for a planned 10-minute trip. However, I have been waiting for 80 minutes due to unforeseen queuing delays. This significantly impacts my schedule, particularly the upcoming meeting with Gloria Campbell at Office. Since there is no congestion in the road network and my travel time from Supermarket to Office is only 10 minutes by car, I can still make the originally scheduled 13:15 departure time if I leave now. Given my profession as a museum curator and the importance of finalizing exhibit details with Gloria, it's essential to maintain this meeting.
Plan revisions: Michael Chen update future activity plan:
[['Office', '14:10', 120, 'drive', 'Ave_2_link_3, St_4_link_1', 'Arrive at Office by 14:20 to meet with Gloria Campbell and finalize exhibit design details.'], ['Gym', '17:45', 60, 'transit', 'Metro_1', 'After work, take Metro_1 from Node_8 to Node_13 and transfer to visit Gym for light training and relaxation.'], ['Uptown apartment', '19:30', 'none', 'transit', 'Metro_2', 'Return home via Metro_2 to Node_1 and spend quiet evening with family before bedtime.']] 

Robert Johnson | Time: Thu 2025-03-13 07:01
I am currently en route to the Gym from Midtown apartment, but there is moderate congestion on St_1_link_2, where David Martinez is already experiencing a 5-minute delay. Although I have not yet encountered this link in my current trip, my planned path includes it, which could impact my punctuality. Given that my work as a fitness instructor requires me to be present early for morning classes, avoiding delays is crucial. Fortunately, alternative routes like using St_2_link_1 and Ave_2_link_2 are available, which bypass the congested area.
Plan revisions: Robert Johnson update current path to ['St_2_link_1', 'Ave_2_link_2'] 

Gloria Campbell | Time: Thu 2025-03-13 07:13
I am currently en route to work on the Factory, departing from Midtown apartment at 07:00. However, I have just encountered severe congestion on St_1_link_1 (Node_4 -> Node_9), with a wait time of 16 minutes reported. This will delay my arrival at the Factory unless I adjust my travel path. My original plan was to take Ave_4 and St_4, which is still viable if I can reroute efficiently to avoid further delays.
Plan revisions: Gloria Campbell update current path to ['Ave_2_link_1', 'Ave_2_link_2', 'St_3_link_1'] 

Christopher Evans | Time: Thu 2025-03-13 08:25
I am currently at the Coffee shop, where I have been discussing client training schedules with Benjamin Taylor. The current traffic state shows no congestion across any links in the network, which is a positive change from earlier delays on Ave_4_link_1 and St_1_link_2. This improved traffic condition allows for more flexibility in my upcoming travel plans. According to recent perception updates, Benjamin will remain at the Gym until at least 08:30 working with clients, so I plan to meet him there by 08:25 as previously coordinated. Since there is no ongoing congestion, I can proceed with confidence along the shortest available path to the Gym.
Plan revisions: Christopher Evans update current path to real-time shortest

Examples of agents gaining new knowledge through reflections on every day travel experiences

Name: Rachel Thompson | Date: Thu 2025-03-13 AM
This week has been a balance of managing the Cinema's oper