Reinforcement Learning as a Meta-Controller for Nav2-MPPI: Scenario-Based Validation in Dynamic Environments

Integrating learning-based components into automated vehicle software stacks creates a key validation challenge: end-to-end learned controllers are hard to verify, while classical planners and controllers depend on manually tuned parameters that may not generalize across operating conditions. The current paper studies an intermediate architecture in which reinforcement learning (RL) does not output motion commands but instead acts as a meta-controller that adjusts the parameters of an established local controller. The system combines a D* Lite global planner with incremental replanning, the Model Predictive Path Integral (MPPI) controller from the ROS 2 Nav2 stack, and a Proximal Policy Optimization (PPO) agent. The agent chooses five MPPI parameters (velocity limits, sampling standard deviations, and an obstacle cost weight) from a bounded action space, using a state composed of goal distance, velocities, cross-track error, and sectorized LiDAR ranges. Because the RL output is restricted to a bounded parameter space and the controller itself is unchanged, the controller’s structure and constraint handling are preserved, and the learned component remains auditable. The approach is validated in Gazebo with a reproducible scenario-generation pipeline that creates matched occupancy grids and simulation worlds of increasing complexity, including static and dynamic variants; in the dynamic variant, an obstacle is injected onto the planned path during execution. Performance is assessed with a unified metric set covering success rate, collision count and duration, time spent in a danger zone, path deviation (mean, max, RMS), detour ratio, execution time, and command smoothness. The hybrid system is compared with A*, D* Lite, and RRT* using default MPPI parameters. In static scenarios, the hybrid system performed comparably to the classical planners, and all except RRT* reached the goal without collision. In dynamic scenarios, the hybrid system completed every run without collision, while D* Lite with default parameters collided once, and the hybrid system reduced execution time by 6-9% relative to D* Lite. Code and results are openly available at jkk-research.github.io/rl_mppi_meta_controller.