Scholarlink

Accurate motion prediction is crucial for human-robot collaboration (HRC), enhancing responsiveness and safety in tasks like object handover. Long Short-Term Memory (LSTM) networks effectively model temporal dependencies in motion data, but their performance is highly dependent on the choice of loss function. This paper compares loss functions based on trajectory and Dynamic Movement Primitive (DMP) errors, utilizing both Mean Squared Error (MSE) and Huber loss. Additionally, we assess VGG-11 and AlexNet as feature extractors to optimize prediction accuracy. Our findings provide insights into improving loss selection and network architectures for real-time collaborative robotics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of Loss Functions for LSTM-Based Motion Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Matija Mavsar,
  • Andraž Čepič,
  • Aleš Ude

摘要

Accurate motion prediction is crucial for human-robot collaboration (HRC), enhancing responsiveness and safety in tasks like object handover. Long Short-Term Memory (LSTM) networks effectively model temporal dependencies in motion data, but their performance is highly dependent on the choice of loss function. This paper compares loss functions based on trajectory and Dynamic Movement Primitive (DMP) errors, utilizing both Mean Squared Error (MSE) and Huber loss. Additionally, we assess VGG-11 and AlexNet as feature extractors to optimize prediction accuracy. Our findings provide insights into improving loss selection and network architectures for real-time collaborative robotics.