GEARing smart environments for pediatric motor rehabilitation

Elena Kokkoni, Effrosyni Mavroudi, Ashkan Zehfroosh, James C. Galloway, Renè Vidal, Jeffrey Heinz, Herbert G. TannerJournal of NeuroEngineering and Rehabilitation 2020 · DOI

Abstract

Preliminary results from this study support the feasibility of both the physical and cyber components of the GEAR system and demonstrate its potential for use in future studies to assess the effects on the co-development of the motor, cognitive, and social systems of very young children with mobility challenges.

Introduction

Traditional pediatric mobility training often focuses on "single-skill" tasks, such as walking on a treadmill. However, infant development is a holistic process where motor, cognitive, and social skills are deeply intertwined. For children with developmental delays, such as those with Down syndrome, the high physical effort required to explore their environment often leads to sedentary behavior, which can further hinder overall development.

The Grounded Early Adaptive Rehabilitation (GEAR) system aims to bridge this gap by providing a smart environment where physical assistance lowers the barrier to movement, while socially-assistive robots provide the motivation necessary to encourage spontaneous exploration and play.

System Architecture and Methodology

The development of the GEAR system followed a structured six-phase approach, moving from the creation of physical components to full system integration and improvement.

Fig. 1: The six-phase development lifecycle of the GEAR system, spanning physical component creation to full system integration.

Physical Components

The GEAR environment consists of a 100 $ft^2$ playground equipped with ramps, stairs, and an Oasus body-weight support (BWS) device—a multi-directional overhead rail system. To encourage social interaction, the system utilizes two robots: NAO, a humanoid for imitation tasks, and Dash, a wheeled robot for chasing games.

Fig. 2: Illustration of the GEAR playground setup, featuring the BWS system, humanoid and wheeled robots, and a network of Kinect sensors.

Cyber Components

A network of five synchronized Kinect sensors monitors the area. The system employs a "cyber-physical" architecture for data processing:

  • Action Classification: The system uses Improved Dense Trajectories (IDT) and Fisher Vector encoding. To manage occlusions from the BWS harness or playground equipment, the researchers utilized a Multiple Instance Learning SVM (MI-SVM). This treats the five camera views as a "bag" of data, allowing the algorithm to exploit complementary information across views.
  • Decision Making: Robot behavior is governed by a Markov Decision Process (MDP). Using "smoothing" techniques to estimate transition probabilities from sparse data, the system allows the robot to adapt its policy (e.g., when to retreat or approach) based on the child's real-time engagement level.

Pilot Study and Results

The study conducted a pilot with three infants (one with Down syndrome, two typically developing) aged 10–24 months over 8 sessions per child.

Key Findings

  • Feasibility and Safety: All subjects completed the protocol with zero adverse events, proving the safety of the BWS-robot integration.
  • Classification Accuracy: The MI-SVM achieved an average accuracy of 72.4%, outperforming standard SVM approaches (68.75%). The system was highly effective at identifying dynamic actions like walking and crawling.
  • Social Interaction: Robots acted as social "actuators." Children frequently adapted their motor actions to interact with the robots, such as assisting the wheeled robot over obstacles. Notably, the child with Down syndrome showed higher visual interest in the robots than typically developing peers.
  • Policy Optimization: AI-driven robot behavior (the "optimal policy") resulted in significantly higher child engagement compared to human-controlled behavior.

Discussion and Future Work

While the pilot demonstrates the viability of the GEAR system, several areas for improvement remain. The current trajectory-based features struggle to distinguish between static poses like sitting and standing; future iterations may incorporate deep learning pose-estimation. Additionally, while the cyber components were validated offline, the next phase involves "closing the loop" to allow the action recognition algorithms to control the robots autonomously in real-time.

Ultimately, the GEAR system represents a promising shift toward behaviorally natural and perceptively complex rehabilitation environments that support the holistic development of young children with mobility challenges.

Summary updated

Abstract

Preliminary results from this study support the feasibility of both the physical and cyber components of the GEAR system and demonstrate its potential for use in future studies to assess the effects on the co-development of the motor, cognitive, and social systems of very young children with mobility challenges.
BibTeX
@article{kokkoni2020gearing,
  author = {Elena Kokkoni and Effrosyni Mavroudi and Ashkan Zehfroosh and James C. Galloway and Renè Vidal and Jeffrey Heinz and Herbert G. Tanner},
  title = {{GEARing smart environments for pediatric motor rehabilitation}},
  journal = {Journal of NeuroEngineering and Rehabilitation},
  year = {2020},
  doi = {10.1186/s12984-020-0647-0},
}
Figure