Safe and Efficient Variational Inference Model Predictive Control with Application to Aggressive Autonomous Driving
Loading...
Author(s)
Yin, Ji
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation investigates Variational Inference Model Predictive Control (VIMPC) as a flexible and scalable framework for trajectory optimization, with a focus on autonomous driving—a domain that demands a delicate balance between high-performance control and strict safety requirements. Existing VIMPC methods, such as Model Predictive Path Integral (MPPI) control, leverage forward simulations using general nonlinear dynamics to reduce the simulation-to-reality gap. However, these approaches often entail high computational costs and lack formal guarantees of safety and robustness, limiting their effectiveness in real-time, safety-critical scenarios. The key research question addressed in this work is: How can we design VIMPC controllers that are computationally efficient and provide formal safety guarantees, while preserving or even enhancing control performance? To this end, the dissertation proposes a set of improvements to the VIMPC framework. These include Covariance Steering to improve sampling efficiency, Conditional Value-at-Risk to incorporate risk-sensitive decision-making, and Control Barrier Functions to enforce formal safety guarantees. Together, these techniques enable real-time, risk-aware control under uncertainty, even with limited computational resources. By bridging theoretical advances in stochastic control, risk metrics, and formal methods with practical considerations for deployment, this research contributes to the development of reliable, high-performance MPC controllers for complex, safety-critical systems.
Sponsor
Date
2025-04-11
Extent
Resource Type
Text
Resource Subtype
Dissertation