Route Planning on Real Terrain

Reinforcement learning agent that finds low-slope routes on real terrain around the Misti volcano.

The Challenge

Finding a walkable route on steep terrain is usually solved with fixed rules. We wanted an agent that learns the route by itself from elevation data, and we wanted to measure how close it gets to the optimal path.

Our Solution

We turned the Copernicus DEM (Digital Elevation Model) GLO-30 of the Misti volcano in Arequipa into a grid environment where steep slopes are penalized. We trained Q-learning, DQN (Deep Q-Network), Double DQN and REINFORCE agents in PyTorch and compared them with Dijkstra.

Key Results

DQN reaches a fixed goal in 100% of test runs
Routes only 7.1% more costly than the optimal Dijkstra path
DQN beats REINFORCE: 100% vs 95% success

Project Info

Client

Cloud4Geo experiment, ready to adapt to your organization

Services

Reinforcement Learning, Terrain Analysis

Year

2026

Interested in similar results?

Let's discuss how we can apply these solutions to your project.