iKnet Model Improvements for Low Computing Resource Environments
Rebuild iKnet models to optimize performance in environments with limited computing resources, focusing on efficiency and accuracy improvements.
Advisor
Team
Sheng-Kai Chen
Project Developer & Paper Revise
Yi-Ling Tsai
Paper Co-Author
Chun-Chih Chang
Paper Co-Author
Yan-Chen Chen
Paper Co-Author
Abstract
As the demand for inverse kinematics (IK) applications in embedded systems grows, developing efficient and accurate IK neural network models for resource-constrained devices remains a critical challenge. In this study, we propose two improved versions of the original IKNet: Improved IKNet and Focused IKNet. These models aim to enhance computational efficiency while maintaining accuracy. We evaluate the three IK models — Original IKNet, Improved IKNet, and Focused IKNet — under different computing environments, including both CPU and GPU settings. The evaluation focuses on three key metrics: memory usage, inference time, and accuracy. Experiment results show that both proposed models significantly outperform the baseline in inference speed and memory efficiency, especially in CPU-only environments, making them well-suited for deployment on embedded systems.