Bayesian Optimization for Learning Nonlinear MPC in Autonomous Agent Navigation
January 1, 2026·
,,,,·
1 min read
Lorenzo Ortolani
G. voss
G. beltrami
F. dorati
T. f. banfi
Abstract
We present a data-efficient framework that learns the cost weights of a nonlinear Model Predictive Control (MPC) navigation stack using Bayesian Optimization. Rather than hand-tuning the controller, we treat closed-loop navigation performance as a black-box objective and optimize the MPC parameters directly from rollouts, balancing tracking accuracy, safety, and smoothness. The approach is validated on an autonomous agent navigation task and deployed on a Unitree Go2 quadruped, yielding robust, generalizable behavior with far fewer trials than grid or manual tuning.
Type
Publication
IEEE ICRA 2026 Workshop
First-author work (with G. Voss, G. Beltrami, F. Dorati, and T. F. Banfi) presented at the IEEE ICRA 2026 Workshop. We learn the weights of a nonlinear MPC navigation controller via Bayesian Optimization, treating closed-loop performance as a black-box objective and deploying the result on a Unitree Go2 quadruped.
Autonomous Agent Navigation
Model Predictive Control
Bayesian Optimization
Legged Robotics
Machine Learning
Authors
Lorenzo Ortolani
(he/him)
Roboticist · Physical AI & VLA for Humanoids
I build Physical AI — robots that perceive, reason, and act in the real world.
My focus is advancing Vision-Language-Action (VLA) policies so that a single
humanoid can learn a diverse library of skills and choose the right one for whatever
industrial task it faces, generalizing far beyond a single demonstration.
I co-founded TalOS Robotics AI to put this in operators’ hands, and I work across the
full stack: locomotion and navigation, learned manipulation, sim-to-real, and the
agent tooling that wraps it all together.
Authors
Authors
Authors
Authors