Title: Safe and Efficient Variational Inference Model Predictive Control with Application to Aggressive Autonomous Driving
Date: Tuesday, April 1st, 2025
Time: 9 AM - 11 AM EST
Location: Virtual
Virtual Link: Zoom
Virtual Meeting ID: 980 237 8481
Ji Yin
Robotics PhD Candidate
Dynamics and Control Systems Laboratory
Georgia Institute of Technology
Committee
Dr. Panagiotis Tsiotras (Advisor), Daniel Guggenheim School of Aerospace Engineering, Georgia Tech
Dr. Samuel Coogan, School of Electrical and Computer Engineering, Georgia Tech
Dr. Ye Zhao, George Woodruff School of Mechanical Engineering, Georgia Tech
Dr. Shreyas Kousik, George W. Woodruff School of Mechanical Engineering, Georgia Tech
Dr. Chuchu Fan, Aeronautics and Astronautics, Massachusetts Institute of Technology
Abstract
Autonomous driving demands a careful balance between optimal performance, safety, and real-time feasibility. Variational Inference Model Predictive Control (VIMPC), particularly Model Predictive Path Integral (MPPI) control, offers flexibility in handling nonlinear dynamics but is hindered by high computational costs, a lack of risk-awareness, and the absence of formal safety guarantees. This research enhances VIMPC by improving sampling efficiency, incorporating risk-aware decision-making, and enforcing formal safety constraints. To optimize trajectory sampling, Covariance-controlled MPPI (CC-MPPI) dynamically adjusts the covariance of its sampling distribution, significantly enhancing efficiency. Risk-aware MPPI (RA-MPPI) integrates Conditional Value-at-Risk (CVaR) to penalize high-risk trajectories, improving decision-making under uncertainty. Shield MPPI enforces safety through Control Barrier Function (CBF) constraints, extending to Belief-space Stochastic MPPI (BSS-MPPI) for robust planning under state estimation uncertainty. Finally, Neural Shield MPPI (NS-MPPI) leverages Neural Control Barrier Functions (NCBF) to further enhance sampling efficiency and overall autonomous driving performance. Validated on 1/28-scale BuzzRacer and 1/5-scale AutoRally platforms, these methods significantly improve efficiency, risk-awareness, and safety, making VIMPC a more viable solution for real-time, high-performance autonomous driving.