Abstract
In recent years, while progress has been made in barrier-free design, the complete elimination of physical barriers such as uneven road surfaces and stairs remains difficult, and wheelchair passengers continue to face significant mobility constraints. This study aims to verify the effectiveness of a transformable quadruped wheelchair that can switch between two modes of movement: walking and wheeled travel. Specifically, reinforcement learning using Proximal Policy Optimization (PPO) was used to acquire walking strategies for uneven terrain and wheeled travel strategies for flat terrain. NVIDIA Isaac Sim was used for simulation. To evaluate the stability of both modes, we performed a frequency analysis of the passenger's acceleration data. As a result, we observed periodic vibrations around 2 Hz in the vertical direction in walking mode, while in wheeled mode, we confirmed extremely small vibrations and stable running. Furthermore, we distilled these two strategies into a single mode-conditional strategy and conducted long-distance running experiments involving mode transformation. The results demonstrated that by adaptively switching between walking and wheeled modes depending on the terrain, mobility efficiency was significantly improved compared to continuous operation in a single mode. This study demonstrates the effectiveness of an approach that involves learning multiple specialized strategies and switching between them as needed to efficiently traverse diverse environments using a transformable robot.