Model the city
Combine the road network with observed parking and charging demand.
Transportation Research Part E · Volume 194 · 2025
Strategic planning of geo-fenced micro-mobility facilities using reinforcement learning
A deep reinforcement-learning framework turns observed urban demand into budget-feasible plans for parking, charging and battery-swapping facilities.
The infrastructure problem
Free-floating mobility is flexible. Its infrastructure is not.
Without designated facilities, vehicles can obstruct public space and create long collection routes. Geo-fenced parking, charging and battery-swapping locations can improve order and accessibility, but cities must decide where to build, which facility type to choose and how much to invest.
2.5M+real-world trips across both city datasets
€500K–€3Msix infrastructure budget levels
3facility types: parking, charging and battery swapping
The Austin shift
Six stations turn a city-wide collection problem into a short route.
In a one-square-kilometre Austin example, 263 vehicles are scattered across the street network. Restricting drop-offs to six strategically placed stations preserves convenient access while dramatically reducing the route required for collection and charging.
161.47 kmcollection route without stations
263vehicles scattered across one square kilometre
2.97 kmcollection route with six stations
>95% · 250 mof vehicles remain within the walking threshold
Planning framework
From observed demand to a budget-feasible infrastructure plan.
Set priorities
Balance service coverage and convenience against capacity, cost and repositioning effort.
Learn placements
A Double Deep Q-Network creates, extends or relocates facilities.
Compare plans
Evaluate candidate plans across budgets and local policy priorities.
Inside the agent
One network, several ways to improve it.
The agent observes demand, then chooses actions that create or move facilities. The decision rule can focus on the largest uncovered demand or on the area reached by a new facility.
Real-world evaluation
The same framework adapts to two very different cities.
Austin's demand is broad and heterogeneous; Louisville's is more centralized. Training on each city's own trips and road network produces locally tailored infrastructure strategies instead of transferring a single placement recipe.
City datasetAustin, Texas
Distributed demand with dense central and southern hotspots.
- Trips
- 2M+
- Road nodes
- 3,240
- Road edges
- 9,149
- Peak parking-coverage gain
- +163%
City datasetLouisville, Kentucky
More centralized demand across a larger road graph.
- Trips
- 500K
- Road nodes
- 4,200
- Road edges
- 12,064
- Peak parking-coverage gain
- +72%
Planning trade-offs
A bigger budget helps, but every objective moves differently.
Across six budgets, additional facilities generally improve coverage and shorten users' distance to the nearest facility. At the same time, a wider infrastructure network can increase collection and redistribution effort. Returns diminish once high-demand areas are saturated.
Service coverageMore parking and charging demand reached
Rises
Distance to a facilityMore convenient access for users
Falls
Operational networkMore locations to collect from and redistribute between
Trade-off
Configurable utility
The planner makes local priorities explicit.
Coverage
Convenience
Capacity shortage
Repositioning effort
Resources
Article and publication record.
Citation
Transportation Research Part E, 194.
Julian Teusch, Bruno Neumann Saavedra, Yannick Oskar Scherr, and Jörg P. Müller. Strategic planning of geo-fenced micro-mobility facilities using reinforcement learning. Transportation Research Part E: Logistics and Transportation Review, 194, 103872, 2025. DOI: 10.1016/j.tre.2024.103872.