MSc Thesis: Stochastic MPC for Autonomous Vehicles in Uncertain Situations

Authors: Qun Zhang, Saeed Salih, Erik Börve
Emails: qunz@chalmers.se, saeedsal@chalmers.se, borerik@chalmers.se
Affiliation: Department of Electrical Engineering, Chalmers University of Technology, Göteborg, Sweden
Organization: Volvo Group

Instruction

• Overwrite carla in the CARLA path with the folder named carla in the file to use the local controller for this project
• Simulations are run via the "main" and "main_EKF" file. This is also where simulations are configured, including e.g., designing traffic scenarios and setting up the optimal controllers.
• Before running, change map to Town06
• Notice: Remember to change the path.

Purpose

The objectives of this thesis are to:

• Construct safety-critical scenarios for a heavy vehicle in the CARLA simulator, emphasizing realistic challenges in autonomous driving.
• Design and implement a Robust Model Predictive Controller (RMPC) that accounts for uncertainties in the ego-vehicle's state and dynamics, ensuring safety and reliability.
• Extend the RMPC to effectively handle uncertainties related to the surrounding vehicles, improving situational awareness and decision-making.

Workflow

Our workflow integrates an Extended Kalman Filter (EKF) with a Model Predictive Controller (MPC) for enhanced accuracy and robustness, depicted in the figures below. Notably, we simulate sensor inputs rather than using actual CARLA sensors to streamline our process.

Workflow Overview:

Vehicle Model:
The vehicle model is shown below:

Trajectory Propagation:
The difference between the simulated vehicle in CARLA and our nominal model is treated as noise. The figure illustrates how this discrepancy propagates over time.

Constraint Definitions:

For detailed constraint definitions, please refer to our supervisor's paper:

E. Börve, N. Murgovski, and L. Laine, "Interaction-Aware Trajectory Prediction and Planning in Dense Highway Traffic using Distributed Model Predictive Control."

If you find the details on constraint definitions helpful or if they've sparked some ideas for your own work, we'd really appreciate it if you could cite our supervisor's paper.

Illustration of the coordinate system

SMPC Theory:
SMPC Constraint Tightening:
To address these issues, we employ Stochastic MPC (SMPC) techniques to tighten state constraints, especially for trailing and lane changing maneuvers.

(a)

(b)

(c)

Figure(a),(b),(c) are the initial constraints and the constraints after the SMPC

Simulation in the CARLA Environment

Simulation in CARLA:

Collision Avoidance Success Rate:  99/100 
(Simulate in CARLA with 100 experiments of randomly generated environment)

Complex Scenerio Simulation in CARLA:

Driving in Heavy Traffic Conditions:

Driving in Heavy Traffic Conditions using EKF:

Comparison of MPC and SMPC: