High-Altitude Balloon Swarms

High-altitude balloons cannot thrust sideways; they steer by changing altitude to ride different wind layers in a stochastic, partially observable flow field. Coordinating a team of them to cover an area is a multi-agent decision problem with severe underactuation and communication limits.

With NRL's Distributed Autonomous Systems Group, we developed multi-agent reinforcement learning methods for distributed area coverage with balloon swarms in complex dynamic flow fields, alongside the simulation environments, path-planning, and optimization tooling to study them (IEEE Aerospace 2026; oral presentation, Big Sky, MT).

This platform is also where the efficiency question becomes existential, since bandwidth and power are scarce in the stratosphere. That constraint motivated Learning When to Act, our communication-efficient RL framework with run-time assurance.