Bayesian Optimization for Learning Nonlinear MPC in Autonomous Agent Navigation
Researchers have developed a new map-free framework for autonomous robot navigation that combines reactive planning with nonlinear Model Predictive Control (MPC). This system uses a LiDAR-based Gaussian occupancy representation and an A* search algorithm to generate collision-free trajectories, which are then tracked by an MPC formulation. To optimize controller parameters, an offline Bayesian optimization scheme utilizing Tree-structured Parzen Estimators (TPE) and a Gaussian Process surrogate was employed. The framework was successfully evaluated on a Unitree Go2 robot in simulation and on the physical hardware, achieving a 90.0% navigation success rate and demonstrating effective parameter transfer from simulation to real-world deployment. AI
IMPACT Enhances autonomous navigation capabilities for mobile robots, potentially improving performance in complex and dynamic environments.