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

Area of Science:

• Robotics
• Artificial Intelligence

Background:

• Robot navigation is evolving towards socially aware strategies for human-robot interaction.
• Reinforcement learning (RL) is advancing socially aware navigation, but reward function design is challenging, especially in congested environments.
• Current manually designed reward functions suffer from hyperparameter issues and poor representation of object characteristics.

Purpose of the Study:

• To address the challenges in designing reward functions for socially aware robot navigation.
• To introduce a novel, adaptable reward function that simplifies tuning and improves performance.
• To enhance the efficiency and effectiveness of deep reinforcement learning (DRL) in human-centric robotic applications.

Main Methods:

• Developed a transformable Gaussian reward function (TGRF) with independently functioning hyperparameters.
• Designed TGRF for adaptability, allowing diverse reward function applications.
• Integrated TGRF within a deep reinforcement learning (DRL) framework for robot navigation tasks.

Main Results:

• TGRF significantly reduces the complexity of reward function tuning.
• The transformable nature of TGRF allows for flexible application across various navigation scenarios.
• Experiments and simulations demonstrated high performance and accelerated learning rates using TGRF in DRL.

Conclusions:

• TGRF offers a more efficient and effective approach to reward function design for socially aware robot navigation.
• The proposed method simplifies the development of robots that can navigate safely and effectively alongside humans.
• TGRF shows promise for advancing the capabilities of DRL in complex, dynamic environments.