Flying Through a Narrow Gap Using End-to-End Deep Reinforcement Learning Augmented With Curriculum Learning and...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Quadrotor navigation in confined spaces, like tilted narrow gaps, presents significant challenges for autonomous systems.
• Existing reinforcement learning (RL) methods struggle with dynamic feasibility and accurate Sim2Real transfer in such scenarios.
• The risk of collision damage limits real-world data collection for training.

Purpose of the Study:

• To develop an end-to-end reinforcement learning framework capable of successfully traversing tilted narrow gaps.
• To address the difficulties in searching for feasible trajectories and the low error tolerance in Sim2Real transfer.
• To achieve gap traversal using deep reinforcement learning without relying on real-world flight data.

Main Methods:

• Utilized curriculum learning to guide the RL agent towards sparse rewards, facilitating the search for dynamically feasible flight trajectories.
• Proposed a novel Sim2Real framework designed to transfer control commands to a real quadrotor effectively, bypassing the need for real flight data.
• Implemented an end-to-end deep reinforcement learning approach for autonomous quadrotor control.

Main Results:

• Successfully demonstrated the ability of the proposed framework to enable quadrotors to traverse tilted narrow gaps.
• Overcame the challenges of sparse rewards and the need for precise trajectory planning in confined environments.
• Achieved robust Sim2Real transfer, allowing control commands trained in simulation to be directly applied to a physical quadrotor.

Conclusions:

• The developed end-to-end reinforcement learning framework provides a viable solution for autonomous quadrotor navigation through tilted narrow gaps.
• This work represents a significant advancement, being the first to accomplish successful gap traversal purely through deep reinforcement learning.
• The proposed methods offer a pathway for more complex autonomous robotic tasks in challenging, constrained environments.