Simulation-to-Reality Hyperparameter Optimization of MPPI Controllers via Bayesian Optimization in NVIDIA Omniverse Isaac Sim

Autonomous mobile robots navigating dynamic environments require robust and efficient control strategies. Model Predictive Path Integral (MPPI) control offers flexibility and computational efficiency but demands precise tuning of multiple hyperparameters, traditionally conducted through heuristic methods. In this paper, we introduce an automated and systematic approach to MPPI hyperparameter optimization using Bayesian Optimization (BO) within a ROS2 and Nav2-based framework. By leveraging a high-fidelity digital twin implemented in NVIDIA Omniverse Isaac Sim, our method automates and accelerates hyperparameter tuning, significantly improving trajectory smoothness, reducing control effort, and improving navigation efficiency. Experimental validation on a real differential drive robot demonstrates strong consistency between simulation-optimized parameters and real-world performance, confirming the effectiveness of our simulation-to-reality approach. This work provides a practical and reproducible method for integrating BO with ROS2 and Nav2, enabling streamlined deployment and adaptive tuning of MPPI controllers in real-world robotics applications.

We present a robust and reproducible framework for tuning Model Predictive Path Integral (MPPI) controllers using Bayesian Optimization (BO) in the ROS2 Navigation2 (Nav2) stack. The approach is based on a high-fidelity digital twin of a differential-drive mobile robot modeled in NVIDIA Omniverse Isaac Sim, enabling realistic and risk-free parameter optimization.

The framework allows automated tuning of core MPPI parameters—including batch size, control noise, and temperature—by minimizing a multi-objective cost function based on trajectory smoothness and goal-reaching efficiency. All experiments are conducted in simulation and directly transferred to the real robot with minimal effort, demonstrating strong sim-to-real consistency.

The simulation-driven BO loop minimizes manual tuning effort and accelerates deployment of MPPI controllers in real-world robotics applications.