Reinforcement Learning for Shared Autonomy in Powered Wheelchair Navigation

Assistive robotics is witnessing a surge in research focusing on designing algorithms and frameworks that offer personalized support to users, considering their intentions and adapting system responses accordingly. This thesis delves into the integration of artificial intelligence in assistive technologies for powered wheelchairs, with a primary emphasis on the challenging problem of shared control through reinforcement learning.

Shared control, also known as shared autonomy, has been extensively studied, especially in the context of powered wheelchairs. Many wheelchair users rely on aid to enhance their everyday autonomy, particularly those who cannot use conventional joystick control interfaces, often encountering frustrations, fatigue, and compromised safety. Existing shared control methods typically involve blending human and autonomous controller decisions or predicting user goals to act autonomously. Unfortunately, such approaches often rely on assumptions like known goal sets, world dynamics models, and user behavior models, which limit adaptability. Motivated by the shortcomings of prior approaches and inspired by recent machine learning advances, this thesis introduces a novel shared control method using deep reinforcement learning within a continuous action space, which lifts the reliance on the aforementioned assumptions.

Initially, a reinforcement learning agent is developed to autonomously navigate complex indoor environments without the need for a map. The agent is trained using a virtual robotic wheelchair and rigorously validated against popular path planning methods. Subsequently, artificial noise is injected into the learned model to simulate disabled user input, enabling the training of an end-to-end shared control system. A modification in the typical reinforcement learning objective ensures compliance with user intentions while simultaneously maximizing future rewards associated with the assistive nature of the system. The shared control system receives noisy user commands and sensor data to generate corrective control commands for the wheelchair. Rigorous simulations and real-world trials with human users demonstrate significant reductions in collisions and increased obstacle clearance, albeit with a trade-off in user satisfaction.

Additionally, this thesis presents a non-intrusive, vision-based head-control interface for powered wheelchairs, employing face detection and head pose estimation. Through human user trials, the effectiveness and performance of this interface are benchmarked, confirming its viability as an alternative to the standard joystick interface. Notably, when combined with the shared control system in further real-world trials, the proposed assistive system proves adept at compensating for the less accurate input of this more challenging interface, resulting in a remarkable 92\% reduction in collisions and improved overall adequacy.

In summary, this thesis introduces a mapless autonomous navigation method for powered wheelchairs, a novel shared control framework employing deep reinforcement learning, and a non-intrusive vision-based head-control interface. The proposed assistive system is empirically validated, showcasing its substantial impact on enhancing user autonomy and safety in powered wheelchairs.