A Causally Inspired Counterfactual Evaluation Framework for Wearable Assistive Robots

Area of Science:

• Robotics
• Human-Robot Interaction
• Biomedical Engineering

Background:

• Evaluating wearable assistive robots in real-world settings is challenging due to the lack of repeatable A/B comparisons.
• Conventional methods assume task consistency between assist-on and assist-off trials, which is often not the case in human-human interactions.
• Caregivers naturally adapt their posture and movement timing during caregiving tasks.

Purpose of the Study:

• To propose a causally inspired diagnostic framework for evaluating wearable assistive robots.
• To address the limitations of conventional A/B comparisons by incorporating movement-fixed counterfactual estimation.
• To interpret assistive responses in complex, real-world caregiving scenarios.

Main Methods:

• Utilized DBN/SCM-inspired time-series modeling and movement-fixed counterfactual estimation.
• Represented multimodal observations including intervention, robot state, movement context, and electromyography (EMG).
• Employed Attention-based Sparse Variational Gaussian Process regressors for node-specific relationship approximation across varying environmental complexities.

Main Results:

• One-step EMG prediction accuracy was insufficient for identifying intervention-sensitive models.
• Controlled validation showed a movement-decoupled model reproducing an EMG-reducing response.
• Real-world case study indicated a near-zero global assist-mediated response (AMR) despite positive pooled A/B differences, with stratified analysis revealing localized supported responses.

Conclusions:

• Context-fixed counterfactual diagnosis can help interpret assistive responses in wearable robots under increasing environmental complexity.
• The proposed framework provides a diagnostic tool for understanding robot assistance beyond simple A/B comparisons.
• Findings highlight the importance of considering movement context for accurate evaluation of assistive robots in dynamic environments.