Reinforcement Learning Aided Robot-Assisted Navigation: A Utility and RRT Two-Stage Approach

Abstract

This work proposes a robot-assisted navigation approach based on user intent adjustment, in the context of robotic walkers. Walkers are prescribed to users with gait disorders so that they can support their body weight on the upper limbs, however, the manipulation of such devices can be cumbersome for some users. Common problems for the users are lack of dexterous upper limb control and visual impairments. These problems can render walkers’ users helpless, making them unable to operate these devices effectively and efficiently. We present a new approach to robot-assisted navigation using a utility decision and safety analysis procedure with user intent adjustments learned by reinforcement learning (RL) and supported on a rapidly-exploring random tree inspired algorithm. The proposed approach offers full control of the assistive platform to the user until obstacles are detected. In dangerous scenarios, corrections are computed in order that the assistive platform avoids collisions and follows social norms, effectively guiding the user through the environment while enforcing safer routes. The experimental validation was carried out in a virtual environment and in a real world scenario using a robotic walker built in our lab (ISR-AIWALKER). Experimental results have shown that the proposed approach provides a reliable solution to the robot-assisted navigation of a robotic walker, in particular the use of utility theory to evaluate candidate motions together with a RL model increases the safety of the user’s navigation.