Related Experiment Video
Updated: Jul 3, 2025
11:01
SSVEP-based Experimental Procedure for Brain-Robot Interaction with Humanoid Robots
Published on: November 24, 2015
13.2K
StairNet: visual recognition of stairs for human-robot locomotion
Andrew Garrett Kurbis1,2, Dmytro Kuzmenko3, Bogdan Ivanyuk-Skulskiy3
1Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. garrett.kurbis@utoronto.ca.
Biomedical Engineering Online
|February 15, 2024
Summary
The StairNet initiative developed deep learning models for robots to recognize stairs using egocentric vision. These models achieve high accuracy and fast inference speeds, aiding robotic prosthetics and exoskeletons.
Area of Science:
- Robotics
- Computer Vision
- Artificial Intelligence
Background:
- Human-robot walking, particularly with prosthetic legs and exoskeletons on complex terrains like stairs, presents a significant challenge.
- Egocentric vision offers a unique advantage in detecting the walking environment, crucial for improving stair navigation.
Purpose of the Study:
- To introduce the StairNet initiative, aimed at fostering the development of deep learning models for visual perception of real-world stair environments.
- To present a comprehensive overview of StairNet's progress, including dataset development and algorithm performance.
Main Methods:
- Development of a large-scale dataset with over 515,000 manually labeled images for stair recognition.
- Evaluation of various deep learning algorithms (2D/3D CNN, CNN-LSTM, ViT), training methods (supervised, semi-supervised), and deployment strategies (mobile, embedded).
Main Results:
- StairNet models achieved high classification accuracy, up to 98.8%, with adaptable trade-offs between accuracy and model size.
- Deep learning models demonstrated fast inference speeds (e.g., 2.8 ms on mobile devices with accelerators) but slower speeds on custom smart glasses (1.5 s).
Conclusions:
- StairNet serves as an effective platform for developing and studying deep learning models for visual perception in human-robot walking environments, focusing on stair recognition.
- This research supports the advancement of next-generation vision-based control systems for mobility assistive technologies like robotic prosthetics and exoskeletons.
