RESUMEN
As the world ages, rehabilitation and assistive devices will play a key role in improving mobility. However, designing controllers for these devices presents several challenges, from varying degrees of impairment to unique adaptation strategies of users. To use computer simulation to address these challenges, simulating human motions is required. Recently, deep reinforcement learning (DRL) has been successfully applied to generate walking motions whose goal is to produce a stable human walking policy. However, from a rehabilitation perspective, it is more important to match the walking policy's ability to that of an impaired person with reduced ability. In this paper, we present the first attempt to investigate the correlation between DRL training parameters with the ability of the generated human walking policy to recover from perturbation. We show that the control policies can produce gait patterns resembling those of humans without perturbation and that varying perturbation parameters during training can create variation in the recovery ability of the human model. We also demonstrate that the control policy can produce similar behaviours when subjected to forces that users may experience while using a balance assistive device.