Sim-to-Real Transfer of Automatic Extinguishing Strategy for Firefighting Robots

Abstract

The automatic extinguishing strategy (AES) is the core of the decision-making system for intelligent firefighting robots. Inspired by the fire extinguishing action of firefighters, designing a vision-based end-to-end AES aligns with human intuition. However, the cost of training agents to learn AES in reality is high. Moreover, training agents in simulation face a gap between simulation and reality, the trained agents often fail in the real world. To solve this problem, we propose a novel AES based on sim-to-real transfer for firefighting robots. This method uses JetGAN, an innovative application of generative adversarial networks (GANs), to translate the simulated jet images into the real domain and uses deep reinforcement learning to construct an AES. First, a genetic algorithm is used to find the simulated jet that closely resembles the input jet image in the real domain, thereby constructing a paired sim-real image dataset. Subsequently, we devise a jet consistency loss and employ the focal frequency loss for JetGAN, which is trained on the paired image dataset. Finally, agents are trained in the simulated environment constructed in Unity3D using jet images translated by JetGAN. The learned AES is capable of transferring to the real world. The experimental results on an actual firefighting robot demonstrate the effectiveness of the proposed sim-to-real transfer. The transferred AES achieved the highest success rate compared with other methods.