Transformable Gaussian Reward Function for Socially Aware Navigation Using Deep Reinforcement Learning.

Robot navigation has transitioned from avoiding static obstacles to adopting socially aware navigation strategies for coexisting with humans. Consequently, socially aware navigation in dynamic, human-centric environments has gained prominence in the field of robotics. One of the methods for socially aware navigation, the reinforcement learning technique, has fostered its advancement. However, defining appropriate reward functions, particularly in congested environments, holds a significant challenge. These reward functions, crucial for guiding robot actions, necessitate intricate human-crafted design due to their complex nature and inability to be set automatically. The multitude of manually designed reward functions contains issues such as hyperparameter redundancy, imbalance, and inadequate representation of unique object characteristics. To address these challenges, we introduce a transformable Gaussian reward function (TGRF). The TGRF possesses two main features. First, it reduces the burden of tuning by utilizing a small number of hyperparameters that function independently. Second, it enables the application of various reward functions through its transformability. Consequently, it exhibits high performance and accelerated learning rates within the deep reinforcement learning (DRL) framework. We also validated the performance of TGRF through simulations and experiments.