Residual dynamics learning for trajectory tracking for multi-rotor aerial vehicles.

Area of Science:

• Robotics
• Control Systems
• Artificial Intelligence

Background:

• High-level planning often uses simplified kinematic models, leading to residual dynamics.
• Accurate modeling of these dynamics is crucial for safe and efficient trajectory tracking, especially in cluttered environments.

Purpose of the Study:

• To develop a technique for modeling residual dynamics between high-level planning and low-level control.
• To improve quadrotor trajectory tracking performance in cluttered and dynamic environments.

Main Methods:

• Utilized a simplified motion model for quadrotor high-level planning.
• Proposed a Sparse Gaussian Process Regression (SGPR) based technique to model residual dynamics.
• Compared SGPR with Data-Driven MPC for residual dynamics error estimation.

Main Results:

• The proposed SGPR technique reduced nominal model error by an average factor of 2 compared to Data-Driven MPC.
• The complete framework outperformed four other trajectory-tracking approaches in collision-free path following.
• Achieved reduced flight time without compromising computational efficiency.

Conclusions:

• The SGPR-based approach effectively models residual dynamics for enhanced quadrotor control.
• This method improves safety and efficiency in complex, obstacle-rich environments.
• The framework offers a computationally efficient solution for aggressive trajectory tracking.