Control Policy via Musculoskeletal Model Simulation

Recently, deep learning-based exoskeleton control has gained significant attention, demonstrating strong potential across a range of applications. However, the developing these controllers critically depends on the training dataset. In deep learning, the quality and diversity of training data largely determine model performance. Yet, datasets collected from human experiments are inherently limited in both scale and task variety due to safety concerns, time constraints, and practical challenges.

To overcome these limitations, our group is pursuing two main objectives:
1. Develop human-like, predictive neuromusculoskeletal simulations by integrating deep reinforcement learning with neuroscience theories and hypotheses, informed by carefully designed human experiments.
2. Leverage these simulations to generate exoskeleton control policies, using domain randomization to minimize the Sim2Real gap.

By achieving these goals, we want to create exoskeleton controllers that not only enhance mobility for diverse populations but also advance rehabilitation, ultimately bridging the gap between simulation and real-world deployment.

Team Members

• Ilseung Park

• Changseob Song