Hummingbird | Ben Morris

Overview

In summer 2022, I worked in Olin College’s RoboLab on Hummingbird, a project exploring swarms of drones in tight indoor environments. The summer goal was to deliver a working proof of concept within ten weeks.

My task was to design and test a path planning algorithm in simulation. Existing approaches like gradient descent, particle swarm optimization, and artificial potential fields struggled in this context because of:

❌ High computation times - drones froze while recalculating paths.
❌ Static environment assumptions - moving drones often caused collisions.
❌ Low precision - most were designed for large open spaces, not centimeter-level accuracy in cramped environments.

So, I created a new algorithm which reduced collisions and enabled real-time path planning in drone swarm simulations.

Solution: Dynamic Gradient Descent (DGD)

To address these issues, I developed a new algorithm, Dynamic Gradient Descent (DGD), inspired by traditional gradient descent but adapted for real-time, dynamic environments.

Instead of precomputing a full path, DGD recalculates each step in fractions of a second:

At each loop, the drone builds a local “topographical map” of the target (valley) and obstacles (mountains).
A metaphorical ball “rolls downhill” for a short duration, providing the next movement direction.
The drone moves one step, polls updated obstacle positions, and repeats.

This spreads computation across the entire flight and allows for continuous course correction without predicting other drones’ paths.

Results

DGD significantly outperformed the other algorithms I tested:

✅ Reduced collisions compared to traditional approaches.
✅ Maintained path efficiency and accuracy.
✅ Minimal computation time enabled real-time control.

Video demo of DGD in action:

In the visualization, lighter colors are “downhill” while darker colors are “uphill.” The black dot represents the drone, dark blue dots represent obstacles (other drones), and the purple star is the target. The drone dynamically navigates obstacles while progressing toward its goal.