Agent Kinematic Domain Randomization for Simulation-Based Performance Evaluation of Wearable Robots

Clinical research has explored the use of wearable robots to aid in functional recovery and locomotion, such as walking and running, for rehabilitation patients. Despite some successful cases, evaluating the performance of wearable robots in real-world environments is both time-consuming and costly. Additionally, while simulation-based approaches are employed to overcome the limitations of real-world testing, challenges remain in transferring the learned control policies of reinforcement learning agents to natural environments. To bridge this “sim-to-real” gap, domain randomization techniques have been applied to the dynamic properties of wearable robots, such as exoskeletons. This approach can also be extended to the dynamics of rehabilitation patient agents. In this study, we investigate the potential of applying physics-based and probabilistic domain randomization methods to enable simulation agents to replicate various disabilities in gait. More importantly, this approach allows for the effective learning of control policies that can be tailored to the specific characteristics of individual patients.