Comparison of Empirical and Reinforcement Learning (RL)-Based Control Based on Proximal Policy Optimization (PPO) for...

Area of Science:

• Biomechanics
• Robotics
• Artificial Intelligence

Background:

• Wearable assistive devices are increasingly used in industrial and medical settings.
• Combining human expertise with artificial intelligence (AI) for personalized assistance is a growing trend.
• AI's potential to surpass human capabilities in customizing support is debated.

Purpose of the Study:

• To investigate the efficacy of AI-driven control strategies for wearable lower-limb assistive devices.
• To compare an empirical control strategy with a Reinforcement Learning (RL) optimized strategy for a thigh exosuit.
• To assess the impact of these controllers on muscle activation and walking performance.

Main Methods:

• The study utilized the Biarticular Thigh Exosuit, which mimics hamstring and rectus femoris muscle actions.
• Two control strategies were tested: an empirical controller and an RL-optimized controller on a neuromuscular model.
• Performance was evaluated by comparing muscle activation (hamstring, gastrocnemius, vastus, rectus femoris) and preferred walking speed in assisted and unassisted modes.

Main Results:

• Both empirical and RL controllers reduced hamstring muscle activation and increased preferred walking speed.
• The empirical controller also decreased gastrocnemius muscle activity.
• The RL-based controller led to increased vastus and rectus femoris muscle activation.

Conclusions:

• Wearable assistive devices can be controlled by both empirical and AI-based methods to aid human walking.
• Reinforcement Learning optimization in assistive device control may not always yield superior results without empirical validation.
• Future AI-driven assistive technologies require careful integration with human expertise and empirical data to ensure optimal and safe performance.