Glider soaring via reinforcement learning in the field.

Area of Science:

• Aerospace Engineering
• Animal Behavior
• Artificial Intelligence

Background:

• Soaring birds utilize atmospheric thermals for navigation and foraging.
• The dynamic and complex nature of thermals presents a significant navigational challenge.
• Understanding bird navigation strategies can inform autonomous flight systems.

Purpose of the Study:

• To develop an autonomous navigation strategy for soaring in atmospheric thermals using reinforcement learning.
• To investigate the environmental cues birds might use for thermal navigation.
• To create a flight policy applicable to autonomous soaring vehicles.

Main Methods:

• A two-meter wingspan glider was equipped with a flight controller for autonomous thermal navigation.
• Reinforcement learning was employed to train the glider based on its field experiences.
• The strategy focused on utilizing onboard estimations of vertical wind accelerations and roll torques.

Main Results:

• A successful autonomous navigation strategy for atmospheric thermals was learned by the glider.
• Vertical wind accelerations and roll torques were identified as key navigational cues.
• The learned policy demonstrated validity through field experiments and simulations.

Conclusions:

• Reinforcement learning can effectively train autonomous systems to navigate complex atmospheric conditions.
• Mechanosensory cues like vertical wind accelerations and roll torques are crucial for soaring navigation.
• The developed strategy has direct applications for autonomous soaring vehicle development.