Implementing Efficient Gas Dispersion Simulation to Train Autonomous Mobile Robots for Gas Distribution Mapping

Reliable gas dispersion simulation is essential for training autonomous mobile robots to perform gas distribution mapping (GDM) using reinforcement learning (RL), but traditional computational fluid dynamics (CFD) solvers are too computationally intensive for RL training that requires thousands of episodes. This work investigates the use of MantaFlow, a visual effects fluid simulation engine, as an efficient alternative for gas dispersion simulation in robotic applications. We integrate MantaFlow into a 3D simulation environment to model ethanol gas dispersion in an indoor environment and deploy a mobile robot equipped with TGS2600 metal oxide gas sensors to collect gas concentration measurements processed using Gaussian Markov Random Field (GMRF). The approach is validated through quantitative comparison of gas concentration maps from three sources: MantaFlow simulation ground truth, simulated robot measurements, and real-world robot experiments. Results demonstrate close correspondence between simulation and reality, with RMSE values of 0.110 (simulation vs. ground truth) and 0.125 (real-world vs. ground truth), representing only a 13.6% difference, validating MantaFlow as a practical and computationally efficient simulation engine for gas distribution mapping applications.