A Causally Inspired Counterfactual Evaluation Framework for Wearable Assistive Robots

Evaluating wearable assistive robots in real-world caregiving is challenging because temporally aligned and repeatable A/B comparisons are rarely available. Conventional evaluations assume that assist-on and assist-off trials are comparable in task content, posture, and movement timing. However, caregivers adjust their posture and timing across human-human interactions and task sequences. This study proposes a causally inspired diagnostic framework based on DBN/SCM-inspired time-series modeling and movement-fixed counterfactual estimation. We represented the multimodal observations using intervention, robot state, movement context, EMG, and context variables. Node-specific relationships were approximated using Attention-based Sparse Variational Gaussian Process regressors. We evaluated the framework at three levels of environmental complexity. These levels comprised controlled trunk flexion, partially controlled bed-to-wheelchair transfer, and real-world caregiving. The proposed framework is intended as a diagnostic counterfactual evaluation tool rather than as a method for strict causal identification. Across experiments, one-step EMG prediction accuracy alone was insufficient to identify intervention-sensitive models. In controlled validation, the selected movement-decoupled robot-only model reproduced an EMG-reducing response consistent with the controlled A/B reference. When fitted to the partially controlled transfer data, the selected structural specification identified an EMG-increasing response in supported contexts. In the real-world caregiving case study, the global assist-mediated response (AMR) was near zero despite a positive pooled A/B difference. However, the stratified analysis identified localized supported responses. The near-zero AMR indicates that the pooled difference was not reproduced through the modeled assist intervention-robot state-EMG pathway under fixed movement context. This result should not be interpreted as evidence of overall device ineffectiveness. These findings suggest that context-fixed counterfactual diagnosis can help interpret assistive responses under increasing environmental complexity.