Skip to main content
Okhare Paul — Portfolio©2026
000
Back

Case study · 01 / 12

Swarm Routing

AI that teaches delivery drones to share the work

Live

Fig. 01 — Swarm RoutingSixty drop-offs, six zones, one swarm. Block a point and watch it think again.
About the project

My MSc dissertation at Birmingham City University: an AI system that routes swarms of delivery drones, tested from 2 to 50 drones.

One drone following a fixed route is easy. Fifty drones sharing a city, with blocked drop-offs, shifting weather and draining batteries, is not. I built the whole system in Python. I generated delivery points, split the city into zones with spectral clustering, planned a route inside each zone, then trained the drones with Q-learning to adapt when things went wrong. The work won the PGXPO 2025 Research Innovation Award and the Faculty Award for Dissertation Excellence.

Main features

What it does

  1. 01Generates realistic delivery points across an urban layout
  2. 02Turns the city into a graph of points and distances
  3. 03Splits the graph into delivery zones with spectral clustering
  4. 04Plans a route inside each zone
  5. 05Assigns a zone to each drone in the swarm
  6. 06Re-routes when a delivery point is blocked
  7. 07Adapts to weather shifts and battery depletion
  8. 08Measures completion rate and delivery time for every run
Approach

How it was solved

  1. 01Started with a nearest-neighbour route as the baseline to beat
  2. 02Partitioned the delivery graph with spectral clustering so drones stop competing for the same points
  3. 03Trained drones with Q-learning over 500 episodes, then scaled to multi-agent reinforcement learning
  4. 04Added disruptions mid-run to test how the swarm recovers
  5. 05Scaled from 2 to 50 drones to prove it holds up
Results

30% higher task completion than static routing, 25% shorter delivery times, and a further 15% gain in completion under disruption.

Technologies

Stack

Language
Python 3.10
Libraries
NumPypandasSciPy (spatial distance)scikit-learnMatplotlib
Routing
Travelling salesman problemNearest-neighbour heuristicDistance matricesGraph adjacency matrixPer-zone re-solving
Clustering
Spectral clusteringPrecomputed affinity graphk-means label assignment
Reinforcement learning
Tabular Q-learningEpsilon-greedy explorationCustom training environmentMulti-agent RL (independent learners)Reward shaping
Simulation
Seeded delivery-point generationBattery and recharge modelNo-fly zonesBlocked-point disruptions
Tools
Google ColabJupyter Notebook

That is the work. Now meet the person who built it.

Next project · 02 / 12Obsiryn