Affordable Access

Learned and Hybrid Strategies for Control and Planning of Highly Automated Vehicles

Authors
  • Bou Ghosn, Agapius
Publication Date
Sep 15, 2023
Source
Hal-Diderot
Keywords
Language
English
License
Unknown
External links

Abstract

The advancement of autonomous vehicles represents a significant leap forward in the pursuit of safer and more reliable modes of transportation. Reaching full autonomy in vehicles has become the central focus of researchers and experts in the field in the last decade. Full autonomy demands precise representation of the vehicle's dynamics across various components within its architecture, to ensure the operation in a wide range of scenarios. To achieve this purpose, this thesis integrates the notion of learning to vehicle observers and planners.Through the integration of hybrid and learned techniques, our objective is to enhance the vehicle's capacity to accurately observe its state. Achieving precise state knowledge is critical as both the planning and control layers rely on this information. We test our observing techniques on real vehicle applications proving the ability of the proposed methods to achieve accurate observations in real-life scenarios, even at the limits of handling of the vehicle. The proposed methods present significant advantages over state-of-the-art methods.After achieving accurate state observations, we propose a simple yet accurate hybrid model in the second stage. This model precisely describes the vehicle's behavior, allowing the development of a planner that can generate feasible trajectories, even in high-dynamic scenarios. An MPPI-based plan and control scheme is proposed and thoroughly tested across various maneuvers. Comparing our approach to the commonly used kinematic bicycle model in planning applications, our results clearly demonstrate the superiority of the proposed method. Notably, the planner utilizing our hybrid model ensures safer and more precise vehicle behavior.This thesis demonstrates the capabilities of the proposed learned and hybrid neural network architectures in accurately representing the complex dynamics of the vehicle. Through simulated and real vehicle experiments, the proposed methods prove their ability to outperform state-of-the-art methods in observing and planning applications.

Report this publication

Statistics

Seen <100 times