A Steering Algorithm for Redirected Walking Using Reinforcement Learning

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Traditional Redirected Walking (RDW) steering algorithms rely on human-engineered logic.
• Recent advancements in reinforcement learning (RL) have shown superior performance in various control tasks.
• There is a need for novel, adaptive steering algorithms in RDW.

Purpose of the Study:

• To investigate the potential of using RL for a novel reactive steering algorithm for RDW.
• To develop a deep neural network-based RL system for RDW steering.
• To compare the performance of the learned RL algorithm against traditional steer-to-center methods.

Main Methods:

• Utilized RL to train a deep neural network for RDW steering.
• The neural network directly outputs rotation, translation, and curvature gains.
• Compared the RL algorithm with steer-to-center using both simulated and real-world paths.

Main Results:

• The RL-based algorithm outperformed the steer-to-center method on simulated paths.
• No significant difference in distance traveled was observed between the algorithms on real paths.
• Demonstrated that RDW is a suitable domain for RL when modeled as a continuous control problem.

Conclusions:

• Reinforcement learning provides a promising framework for developing optimal Redirected Walking steering algorithms.
• The proposed RL approach offers a novel and effective alternative to traditional RDW steering.
• Further research can build upon this framework to enhance RDW user experiences.