Acquisition and Evaluation of Walking and Wheel Movement for a Transformable Quadruped Wheelchair Using Reinforcement Learning

Despite advancements in accessibility, completely eliminating obstacles like stairs and steps remains challenging, presenting significant mobility issues for wheelchair users. This study investigates a transformable quadruped wheelchair capable of switching between walking and wheeled modes. Reinforcement learning and frequency analysis were used to evaluate the effectiveness of four-leg walking on complex terrain and wheeled movement on flat terrain with the user on board. ISAAC SIM was employed for simulation, while proximal policy optimization (PPO) served as the reinforcement learning algorithm. For walking mode, parallel simulation environments with six types of terrains and 2,048 wheelchairs were utilized for training. In contrast, wheeled movement was learned for flat terrain. Acceleration data were collected to evaluate vibrations in both modes using frequency analysis. The results showed notable periodic vibrations around 2 Hz on the Z-axis during walking mode, while wheeled mode exhibited minimal vibrations across all axes, ensuring smoother movement. These findings suggest the desirability of adaptive switching between walking and wheeled modes depending on the environment. Future work includes optimizing automatic mode transitions.