Related Experiment Video
Updated: Sep 22, 2025
06:49
Evaluation of a Smartphone-based Human Activity Recognition System in a Daily Living Environment
Published on: December 11, 2015
9.0K
Unsupervised Sim-to-Real Adaptation for Environmental Recognition in Assistive Walking
Summary
This study introduces an unsupervised method to improve terrain recognition for prosthetic legs, bridging the gap between simulated and real-world data. This enhances prosthetic mobility and adaptability for amputees in diverse environments.
Area of Science:
- Robotics
- Artificial Intelligence
- Biomedical Engineering
Background:
- Powered lower-limb prostheses utilize vision sensors for environmental recognition to restore amputee mobility.
- The sim-to-real gap, caused by differences in real-world terrains and sensor limitations, hinders supervised learning for prosthetic adaptation.
- Existing methods struggle to achieve accurate terrain classification in diverse, unstructured environments.
Purpose of the Study:
- To develop an unsupervised sim-to-real adaptation method for accurate environmental classification in powered lower-limb prostheses.
- To enhance terrain-adaptive locomotion for amputees by overcoming limitations of simulated data.
- To enable robust performance in real-world conditions without requiring labeled real-world data.
Main Methods:
- Augmented simulated environments were generated from a virtual camera perspective to mimic real-world conditions.
- Unsupervised domain adaptation was employed, training a network on simulated and unlabeled real-world data to minimize domain shift.
- The adaptation network, comprising a feature extractor and two classifiers, was trained to classify five common terrains: level ground, stair ascent/descent, and ramp ascent/descent.
Main Results:
- The unsupervised method achieved high average accuracies of 98.06% (indoor) and 95.91% (outdoor) across eight subjects.
- These results closely approach the performance of supervised learning methods that use labeled real-world data (98.37% indoor, 97.05% outdoor).
- Visualizations demonstrated the network's ability to extract and interpret essential terrain features.
Conclusions:
- The proposed unsupervised sim-to-real adaptation method effectively bridges the reality gap for prosthetic environmental classification.
- This approach enables accurate real-world terrain recognition and successful sim-to-real transfer, crucial for adaptive prosthetic locomotion.
- The findings show significant potential for improving prosthetic functionality and amputee mobility in complex environments.
