From Pilot Data to Protocol: Sample-Size Guidance for Multimodal Intent Detection in Assistive Wearable Robotics

Designing a sensible data-collection protocol for intent detection in assistive exoskeletons and wearable robots is challenging: too little data yields unstable models, while collecting “as much data as possible” is costly and often unnecessary. This challenge is particularly relevant in social robotics and human-robot interaction, where human activity recognition (HAR) from biosignals such as EMG is used to infer user intent and physical state for safer, more adaptive assistance. We propose a pilot-to-protocol procedure that turns a small pilot recording into concrete sample-size guidance using learning curves; the methodology is general across model families and sensor modalities, though the specific numerical thresholds it produces are inherently pilot-specific. Using a 32-channel high-density EMG (HD-EMG) grid on the thigh and a co-located IMU, we record eight trials of six lower-limb activities from a
single healthy participant and extract 100 ms windows. We compare RF, SVM, LDA, and a ResNet-18 CNN under leave-one-trial-out (LOTO) evaluation, estimating the training fraction needed to reach 90 % of each model’s peak accuracy and the plateau where adding 10 % more data yields < 1 percentage-point gain. On this single-subject pilot dataset, RF/SVM/LDA typically meet both criteria after 20–30 % of available windows, whereas the CNN continues to improve up to approximately
70–90 %; these values should be interpreted as planning targets for this specific recording, not as universal thresholds. IMU features outperform HD-EMG alone, and EMG+IMU fusion achieves the highest accuracy.
Overall, the protocol provides per-model sample targets and a princi-
pled stopping rule to reduce recording and recalibration burden in data-
efficient exoskeleton intent detection for human augmentation and well-
being.