A Novel Learning-Based Trajectory Generation Strategy for a Quadrotor.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Quadrotor navigation traditionally relies on complex, pre-programmed paths.
• Real-time decision-making for dynamic environments remains a challenge.
• Integrating human-like behaviors can enhance navigation efficiency and reliability.

Purpose of the Study:

• To propose a learning-based trajectory generation framework for quadrotors.
• To achieve real-time, efficient, and practice-reliable navigation.
• To enable human-like decision-making for improved quadrotor control.

Main Methods:

• A framework combining reinforcement learning (RL) and imitation learning (IL) for trajectory generation.
• A real-time local planner integrating learning and optimization techniques.
• An RL policy designed to provide human-like commands for temporal and spatial optimality.
• Convex optimization for real-time trajectory calculation based on RL policy decisions.

Main Results:

• The proposed framework generates smooth, flexible, and efficient quadrotor trajectories in real-time.
• Human-like decision-making via RL policy enhances navigation performance.
• Imitation learning improves generalization and accelerates training.
• Comparative experiments demonstrate superior efficiency, smoothness, and flexibility.

Conclusions:

• The developed framework offers a practice-oriented, intelligent trajectory generation solution for quadrotors.
• Integration of human-like decision-making and model-based optimization yields high-quality trajectories.
• The approach shows significant advantages over existing methods in diverse environments.