Reinforcement-Learning-Based Assistance Reduces Squat Effort with a Modular Hip--Knee Exoskeleton

A team of researchers has demonstrated that an AI-powered exoskeleton can significantly reduce the physical effort of squatting. In a study published on arXiv, Neethan Ratnakumar, Mariya Huzaifa Tohfafarosh, Saanya Jauhri, and Xianlian Zhou developed a neural network controller trained via reinforcement learning (RL) in a physics-based simulation. This controller was deployed on a modular exoskeleton that assists both the hip and knee joints.

Technically, the system works by using the controller to generate real-time assistance torques based on a user's recent joint-angle and velocity history. In trials with five healthy adults performing repetitive squats, the active 'Assistance' condition was compared against a 'Zero-Torque' exoskeleton mode and a 'No-Exo' baseline. The key result was a reduction in net metabolic rate—a measure of physiological effort—by about 10% when using the AI-assisted exoskeleton. The controller successfully adapted its torque output to each subject's unique kinematics, creating a personalized assistance profile.

The context for this work is the high physical demand of squatting in industries like manufacturing and logistics, where repetitive lifting is common. While the study showed promising metabolic savings, it also noted a trade-off: assisted trials resulted in slightly reduced squat depth. This highlights an area for future refinement in balancing assistance with full range of motion. The research represents a step toward intelligent, wearable robots that can learn to augment human strength and endurance in physically demanding jobs, potentially reducing fatigue and injury risk.