The problem
Human activity recognition with energy-harvesting body-area networks sounds great on paper — wrist, ankle, chest sensors all reporting, no batteries — but in practice each node is independently intermittent. Some sensors miss inferences while others overproduce. The system has to coordinate them without a central, always-on clock.
What Origin does
An activity-aware scheduler decides which sensor attempts the next inference based on (a) the energy the sensor expects to harvest in the next window and (b) which sensor’s vantage point is most useful for the predicted activity class.
A lightweight ensemble aggregator fuses partial and confident results from whichever nodes successfully completed their slice. The ensemble is structurally tolerant to missing voters — losing one sensor to a bad harvest window degrades gracefully instead of catastrophically.
Results
2.5–5% higher classification accuracy than standard single-sensor HAR baselines under sporadic harvested power, with negligible coordination overhead. The ensemble structure also dramatically increases the number of completed inferences per minute — the metric that actually matters for practical wearable HAR.
Why it matters
This was an early demonstration that coordination, not bigger models, is the right axis for improving HAR on energy-harvesting bodies. The DATE community recognized it: Origin was a Best Paper Nominee at DATE ‘21.