Champion-level drone racing using deep reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• First-person view (FPV) drone racing requires high-speed piloting and precise navigation.
• Autonomous drones face challenges in operating at physical limits using only onboard sensors.

Purpose of the Study:

• To develop an autonomous system capable of competing at the level of human world champions in FPV drone racing.
• To demonstrate the feasibility of advanced AI in high-speed, sensor-limited robotic navigation.

Main Methods:

• The Swift system integrates deep reinforcement learning (RL) trained in simulation.
• Real-world flight data was incorporated to enhance the RL model's performance.
• The autonomous system was tested in head-to-head races against professional human pilots.

Main Results:

• Swift demonstrated competitive performance against human world champions in real-world races.
• The autonomous system achieved the fastest recorded race time.
• Swift won multiple races against elite human competitors.

Conclusions:

• Swift represents a significant advancement in autonomous mobile robotics and machine intelligence.
• Hybrid learning approaches combining simulation and real-world data are effective for complex robotic tasks.
• This research paves the way for deploying advanced AI in other dynamic physical systems.