Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Load Estimation for Load-lifting Exoskeletons Using Insole Pressure Sensors and Machine Learning

Kaida Wu, Peihao Xiang, Chaohao Lin, Ou Bai

TL;DR

This work addresses real-time load estimation for industrial upper-limb exoskeletons by leveraging insole pressure sensors as a non-invasive, posture-insensitive modality. It explores two data representations—channel-based pressure vectors and plantar-pressure maps—and evaluates three regression models (Elastic Net, SVR, MLP) alongside MobileNetV2 with two transfer-learning strategies for the map input. The channel-based approach with SVR achieves robust cross-subject generalization, attaining a mean MAE of $0.547\,kg$ over loads from $2$ to $10$ kg, with unseen-load interpolation demonstrating practical applicability; map-based methods show potential under high loads but require broader validation. Overall, the results support insole sensing as a viable, low-friction pathway for real-time, adaptive exoskeleton control, paving the way for future work on temporal modeling and closed-loop integration.

Abstract

To enhance lifting-load estimation accuracy in industrial upper-limb assistive exoskeletons, this study proposes a machine learning-based approach using insole pressure sensors. Unlike traditional methods that rely on electromyography (EMG), force sensors, or posture data, insole pressure sensors provide a non-invasive, posture-independent, and stable solution suitable for long-term use. Lifting load data ranging from 2 to 10 kg (0.5 kg intervals) were collected from five subjects. Two data representations were investigated: channel-based vectors and map-based images. For the channel-based approach, conventional regression models (SVR, MLP, and Elastic Net) were trained on pooled data from all subjects to assess inter-subject generalization, specifically testing the ability to infer load levels unseen during training. In parallel, a preliminary feasibility study was conducted for the map-based deep learning model (MobileNetV2) using inner-subject data. Results indicate that the channel-based SVR achieved the most balanced accuracy and generalization performance, with a mean absolute error of 0.547 kg. These findings demonstrate the feasibility and advantages of using insole pressure data for variable load estimation, supporting control strategies in industrial exoskeleton applications.

Real-Time Load Estimation for Load-lifting Exoskeletons Using Insole Pressure Sensors and Machine Learning

TL;DR

This work addresses real-time load estimation for industrial upper-limb exoskeletons by leveraging insole pressure sensors as a non-invasive, posture-insensitive modality. It explores two data representations—channel-based pressure vectors and plantar-pressure maps—and evaluates three regression models (Elastic Net, SVR, MLP) alongside MobileNetV2 with two transfer-learning strategies for the map input. The channel-based approach with SVR achieves robust cross-subject generalization, attaining a mean MAE of over loads from to kg, with unseen-load interpolation demonstrating practical applicability; map-based methods show potential under high loads but require broader validation. Overall, the results support insole sensing as a viable, low-friction pathway for real-time, adaptive exoskeleton control, paving the way for future work on temporal modeling and closed-loop integration.

Abstract

To enhance lifting-load estimation accuracy in industrial upper-limb assistive exoskeletons, this study proposes a machine learning-based approach using insole pressure sensors. Unlike traditional methods that rely on electromyography (EMG), force sensors, or posture data, insole pressure sensors provide a non-invasive, posture-independent, and stable solution suitable for long-term use. Lifting load data ranging from 2 to 10 kg (0.5 kg intervals) were collected from five subjects. Two data representations were investigated: channel-based vectors and map-based images. For the channel-based approach, conventional regression models (SVR, MLP, and Elastic Net) were trained on pooled data from all subjects to assess inter-subject generalization, specifically testing the ability to infer load levels unseen during training. In parallel, a preliminary feasibility study was conducted for the map-based deep learning model (MobileNetV2) using inner-subject data. Results indicate that the channel-based SVR achieved the most balanced accuracy and generalization performance, with a mean absolute error of 0.547 kg. These findings demonstrate the feasibility and advantages of using insole pressure data for variable load estimation, supporting control strategies in industrial exoskeleton applications.

Paper Structure

This paper contains 12 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Implementation and adaptation of insole sensors
  5. The data collection and experiment description
  6. Data Preprocessing
  7. Methods for load estimation
  8. Experimental Hyper-parameter Settings
  9. Results and Discussion
  10. Results
  11. Discussion
  12. CONCLUSIONS

Figures (3)

  • Figure 1: (a) The overall layers of adapted insole sensor. (b) The insole thermal-map of a real pressure data
  • Figure 2: Insole sensor data collection from 2 kg to 10 kg loads
  • Figure 3: (A) MAE distribution of Channel-based approaches, which are SVR, MLP, and Elastic Net, across five subjects (subject 1 to 5) using insole pressure data. The load range is 2 kg to 10 kg. (B) MAE distribution for unseen load levels (3 kg, 6 kg, 9 kg) to evaluate models inference capability . (C) Performance comparison of channel-based approach ( SVR ) and map-based approach ( MobileNetV2 with Linear Probing (LP) and Full Fine-Tuning (Full-FT))on load estimation. (D) Under the map-based approach, MAE distribution for unseen load levels (3 kg, 6 kg, 9 kg) to evaluate models inference capability.