Autonomous navigation of stratospheric balloons using reinforcement learning.

Area of Science:

• Artificial Intelligence
• Aerospace Engineering
• Autonomous Systems

Background:

• Stratospheric balloon navigation faces challenges like wind variability, forecast errors, and sparse data.
• Conventional control methods are insufficient for real-time decision-making in dynamic stratospheric environments.

Purpose of the Study:

• To develop a high-performing flight controller for superpressure balloons using reinforcement learning.
• To address the challenge of reinforcement learning with imperfect data in physical systems.

Main Methods:

• Utilized reinforcement learning (RL) with data augmentation and a self-correcting design.
• Deployed the RL controller on Loon superpressure balloons globally.
• Conducted a 39-day controlled experiment over the Pacific Ocean.

Main Results:

• The RL controller demonstrated superior performance compared to Loon's previous algorithm.
• The controller proved robust against diverse stratospheric wind conditions.
• Successfully overcame the obstacle of applying RL to imperfect real-world data.

Conclusions:

• Reinforcement learning offers an effective solution for autonomous control in complex, real-world scenarios.
• This approach is suitable for systems requiring continuous interaction with dynamic environments.
• Paves the way for advanced AI agents in autonomous navigation and control.