Inverse RL Scene Dynamics Learning for Nonlinear Predictive Control in Autonomous Vehicles.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Autonomous navigation systems require robust trajectory prediction and control.
• Model Predictive Control (MPC) is a common control strategy, but often relies on simplified models.
• Integrating real-time environmental understanding into control is crucial for safety and efficiency.

Purpose of the Study:

• To introduce a novel deep learning-based nonlinear model predictive controller with scene dynamics (DL-NMPC-SD).
• To enhance autonomous navigation by learning scene dynamics for improved trajectory estimation and system modeling.
• To evaluate the performance of DL-NMPC-SD against established and state-of-the-art methods.

Main Methods:

• Encoding a scene dynamics model within a deep neural network using temporal range sensing data.
• Utilizing an Augmented Memory component to integrate range-sensing observations and system states.
• Employing inverse reinforcement learning (IRL) and a modified deep Q-learning (DQL) algorithm for controller training.

Main Results:

• DL-NMPC-SD demonstrated effective trajectory estimation and system model adjustment.
• The method showed competitive or superior performance compared to Dynamic Window Approach (DWA), End2End, and RL methods.
• Successful validation across simulation, indoor/outdoor mobile robot platforms, and full-scale autonomous driving tests.

Conclusions:

• DL-NMPC-SD offers a promising approach for advanced autonomous navigation by integrating learned scene dynamics into model predictive control.
• The deep learning framework effectively approximates complex operating conditions and enhances predictive capabilities.
• The controller's adaptability and performance across diverse environments highlight its potential for real-world applications.