Curriculum-Based Reinforcement Learning for Autonomous UAV Navigation in Unknown Curved Tubular Conduits.

Area of Science:

• Robotics and Artificial Intelligence
• Autonomous Systems
• Sensor Fusion

Background:

• Autonomous navigation in confined tubular environments presents significant challenges due to geometric constraints and limited perception.
• Existing methods often rely on prior geometric knowledge or explicit centerline information, creating an information asymmetry.

Purpose of the Study:

• To develop and evaluate a reinforcement learning (RL) approach for autonomous drone navigation in unknown 3D tubular environments.
• To enable navigation using only local Light Detection and Ranging (LiDAR) and conditional visual data, compensating for the absence of a geometric model.

Main Methods:

• A Proximal Policy Optimization (PPO) based RL agent trained with curriculum learning on increasingly complex geometries.
• A turning-negotiation mechanism integrating direct visibility, directional memory, and LiDAR symmetry cues for stable navigation under partial observability.
• Comparison against a deterministic Pure Pursuit algorithm baseline with explicit centerline access.

Main Results:

• The RL agent demonstrated robust and generalizable navigation capabilities, consistently outperforming the deterministic controller.
• The turning-negotiation mechanism was crucial for stable navigation in scenarios with frequent loss of visual centerline.
• Learned behaviors transferred effectively to a high-fidelity 3D environment with continuous physical dynamics.

Conclusions:

• The proposed RL framework provides a complete solution for autonomous navigation in unknown tubular environments, addressing the challenge of turn negotiation.
• The approach offers a viable solution for applications in industrial, underground, and medical fields requiring navigation through narrow, weakly perceptive conduits.