Rollover-Aware Data-Driven Lateral Control for Tractor-Trailer Evasive Maneuvers

Abstract

In this dissertation, a new rollover-aware evasive maneuver algorithm is developed. Specifically, a stochastic nonlinear model predictive controller is created, which can successfully perform an emergency maneuver in the desired distance while also satisfying the rollover prevention constraints. Many active safety systems for tractor-trailer rollover prevention are reactive systems in the sense that they wait for the vehicle to exceed certain thresholds before applying a corrective measure to ensure safety. This work aims to develop an evasive maneuvering algorithm that explicitly accounts for safety thresholds satisfying both obstacle avoidance and rollover avoidance objectives. Experimental data from an emergency braking maneuver is used to set expected performance standards for an evasive maneuver algorithm. A parameter-free controller is developed that can operate even in the presence of system delays. Both modeling and estimator uncertainty are considered so that intelligent methods of robust constraint enforcement can be implemented.

Typically, parametric models are used for the development of a tractor-trailer lateral controller; however, due to the large number of parameters that can change when trailers are exchanged and the associated difficulty of estimating the new parameters, a parameter-free model is developed. The parameter-free model, called the ultra-local model, utilizes only measurements of the controlled state to form an accurate local model of the dynamic system. This ultra-local model is developed in such a way that even if time delays exist in the system or in the measurements themselves, stable estimation of the ultra-local model can still occur, which is a novel development. The single tunable parameter in the ultra-local model is then parameterized as a function of the vehicle velocity to allow for updating during an evasive maneuver when estimating the parameter may not be possible.

The ultra-local model is tested within the context of a nonlinear model predictive controller. While this implementation successfully performed obstacle avoidance, it was unable to enforce the acceleration constraints, which enforce rollover prevention, with any regularity. To fix this, the sources of uncertainty in the system were analyzed, and the Pontryagin difference, along with chance-constraints, were utilized to generate a robust constraint set. These robust constraints were then implemented in a Stochastic Nonlinear Model Predictive Control (SNMPC) formulation.

Results from the SNMPC simulations demonstrate that both obstacle avoidance and rollover prevention are feasible within the context of an evasive maneuver scenario. When loaded to 29483kg and driven at 22.3m/s the tractor-trailer was able to avoid an obstacle 50m away, which is the distance it takes a vehicle with these parameters to stop with the aid of ABS. Furthermore, even with the influence of sensor noise, the rollover prevention acceleration constraint was satisfied more than 99% of the time, which is the specified threshold set by the chance-constraint. This proves the feasibility of the developed methods for use as a new form of rollover-aware evasive maneuver algorithm.