Abstract
Autonomous robots have the potential to serve as versatile caregivers thatimprove quality of life for millions of people worldwide. Yet, conductingresearch in this area presents numerous challenges, including the risks ofphysical interaction between people and robots. Physics simulations have beenused to optimize and train robots for physical assistance, but have typicallyfocused on a single task. In this paper, we present Assistive Gym, an opensource physics simulation framework for assistive robots that models multipletasks. It includes six simulated environments in which a robotic manipulatorcan attempt to assist a person with activities of daily living (ADLs): itchscratching, drinking, feeding, body manipulation, dressing, and bathing.Assistive Gym models a person's physical capabilities and preferences forassistance, which are used to provide a reward function. We present baselinepolicies trained using reinforcement learning for four different commercialrobots in the six environments. We demonstrate that modeling human motionresults in better assistance and we compare the performance of differentrobots. Overall, we show that Assistive Gym is a promising tool for assistiverobotics research.