Motion Adaptation Across Users and Tasks for Exoskeletons via Meta-Learning

2025-09-17

AI Summary

Addressing the challenge of rapid adaptation to novel users and unseen tasks in exoskeleton systems, this paper proposes a Model-Agnostic Meta-Learning (MAML)-based meta-imitation learning framework. The method integrates simulation-driven elbow joint kinematic modeling, multi-source keypoint extraction and motion retargeting from RGB video and motion-capture data, task-specific neural network prediction, and gravity-compensated PD control to generate personalized reference trajectories. Its key innovation lies in introducing meta-learning for exoskeleton personalization, enabling cross-user and cross-task adaptation within a single, few-shot interaction. Experimental results demonstrate that, for new users performing untrained tasks, muscle activation decreases by 23.6% and metabolic energy consumption drops by 18.4%, significantly outperforming conventional approaches. These findings validate the framework’s superior generalizability, adaptation efficiency, and real-world effectiveness.

Technology Category

• Machine Learning: Imitation Learning & Inverse Reinforcement Learning
• Intelligent Robots: Manipulation
• Multiagent Systems: Adversarial Agents

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Wearable exoskeletons can augment human strength and reduce muscle fatigue during specific tasks. However, developing personalized and task-generalizable assistance algorithms remains a critical challenge. To address this, a meta-imitation learning approach is proposed. This approach leverages a task-specific neural network to predict human elbow joint movements, enabling effective assistance while enhancing generalization to new scenarios. To accelerate data collection, full-body keypoint motions are extracted from publicly available RGB video and motion-capture datasets across multiple tasks, and subsequently retargeted in simulation. Elbow flexion trajectories generated in simulation are then used to train the task-specific neural network within the model-agnostic meta-learning (MAML) framework, which allows the network to rapidly adapt to novel tasks and unseen users with only a few gradient updates. The adapted network outputs personalized references tracked by a gravity-compensated PD controller to ensure stable assistance. Experimental results demonstrate that the exoskeleton significantly reduces both muscle activation and metabolic cost for new users performing untrained tasks, compared to performing without exoskeleton assistance. These findings suggest that the proposed framework effectively improves task generalization and user adaptability for wearable exoskeleton systems.

Problem

• Developing personalized exoskeleton assistance algorithms for task generalization
• Adapting motion control across new users and untrained tasks efficiently
• Reducing muscle activation and metabolic cost during exoskeleton-assisted activities

Innovation

• Meta-imitation learning for motion prediction
• RGB video data retargeting for simulation training
• MAML framework for rapid user-task adaptation