Trajectory Generation by Chance-Constrained Nonlinear MPC With Probabilistic Prediction.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• High-quality trajectory generation often overlooks uncertain future movements of dynamic obstacles.
• Existing methods struggle to effectively predict obstacle trajectories within a defined prediction horizon.

Purpose of the Study:

• To develop a robust trajectory generation framework that accounts for uncertain future obstacle positions.
• To enhance collision avoidance by incorporating probabilistic predictions of moving obstacles.

Main Methods:

• Utilized a variational Bayesian Gaussian mixture model (vBGMM) to predict future obstacle trajectories and associated uncertainty.
• Formulated uncertain future trajectories as collision regions using confidence ellipsoids.
• Implemented chance constraints within a nonlinear model predictive control (NMPC) framework to manage collision probabilities.

Main Results:

• The vBGMM framework effectively predicted future obstacle positions and their uncertainty.
• Probabilistic prediction enabled earlier collision avoidance compared to methods without prediction.
• Generated trajectories exhibited reduced tracking errors and maintained greater distances from obstacles.

Conclusions:

• The proposed approach enhances trajectory generation safety and efficiency in dynamic environments with uncertain obstacles.
• Integrating probabilistic predictions into NMPC provides a more reliable collision avoidance strategy.
• This methodology offers significant improvements over traditional trajectory planning methods lacking predictive capabilities.