MagPie, a protocol built around a microampere wake-up radio and a separately backed low-power real-time clock, completed all 100 first-rendezvous trials in every reported charging scenario. A first rendezvous is the initial point at which two devices find each other and synchronize their timing. In the same tests, Find completed all trials in Office, Stairs and Washer, but only 93 in Jogging and 36 in Cars. Find's restricted mean synchronization time, a measure that includes the study's preset cutoff, was 82.7 to 755.8 times MagPie's.
The design is intended to widen the window for first contact and retain the acquired timing phase for later exchanges. Its wake-up radio listens for a trigger at microampere scale, while the separately backed clock preserves the timing reference. The collection procedure uses component discovery to build a versioned timing tree, then assigns every reporter a unique slot.
First contact was the clearest gain
The pairwise simulation used 100 independent first-rendezvous trials for each setting. They were arranged as ten trials in each of ten seed blocks and administratively censored at 5,000 seconds. The wider evaluation kept three questions separate: pairwise synchronization, a one-collision-domain slotted-Aloha model in which devices share the same transmission space, and controlled STM32WL33 hardware experiments.
After discovery, both protocols completed 100 exchanges in the first four scenarios, but MagPie's mean sustained interval was 1.8 to 3.6 times shorter than Bonito's. Bonito did not discover a peer in Cars before the study horizon. The scheduled-retry analysis is a conditional bound, not a deterministic guarantee: if its coverage premise fails, the tail guarantee no longer applies.
Energy and crowding still set the limits
In the collision-domain model, the role modulus was adapted to the number of eligible contenders. At a tested density of 120 components, that setting required a mean 318.0 formation opportunities, with a 95% confidence interval of 316.1 to 320.0. The 95th-percentile result was 359 opportunities, with a bootstrap interval of 353.1 to 362.1. A fixed role modulus of 32 pushed the 95th percentile to 725 opportunities. These counts are opportunities to try, not wall-clock formation times, because energy readiness was omitted.
The network energy model was more favorable in stable charging patterns than in variable ones. Receiver-side delivery was 99.7% to 99.8% in Stairs and Office, 89.0% in Washer, and about 70% in Jogging and Cars. Because the modeled channel was error-free, the reported losses were energy-related by construction.
At 120 reporters, modeled collection time was 3.57 seconds in Office, Stairs and Washer, 4.32 seconds in Jogging and 11.46 seconds in Cars. Mean receiver-side delivery at that size was 99.8%, 99.7%, 88.3%, 70.4% and 70.2% in the same order. The scalability result does not address collisions, interference or an intermittent sink, so it applies only to the conditions included in the model.
Stress tests exposed weak points
The timing system was also tested against forced independent clock loss on every cycle. With no forced loss, modeled delivery was 99.8% and there were no re-alignments. At a per-cycle loss probability of 0.2, delivery was 80.6%, 79.2%, 80.2% and 79.5% for networks of 5, 10, 20 and 40 reporters. Mean re-alignments rose from 52.5 at 5 reporters to 111.8, 215.4 and 448.9 as the network grew. This was an imposed independent stress model, not a measured hardware failure distribution.
Timing estimates were another fault line. With no artificial cycle-estimation error, joint readiness ranged from 84.7% to 94.1%, above the stated 80% lower bound under the study's 0.9 coverage setting. Under a 25% underestimate, readiness was 25.4% to 74.5%, while a 25% overestimate lengthened every scheduled cycle by about 25%.
Hardware results were deliberately narrower
On hardware, controlled charging times between 500 and 5,000 milliseconds produced 459 complete latency pulses for MagPie and 40 for Find. The normalized alignment rates were 478 and 43 per hour, respectively. Mean latency was 3,678 milliseconds for MagPie versus 24,012 for Find, while the maximums were 34,405 and 222,942 milliseconds. These traces show protocol execution under controlled conditions, not ambient harvesting in the field.
In a separate network run, five reporting boards and one sink operated for 11.05 hours. The network formed in 945 seconds, and the sink logged 2,998 packets with a mean source-to-sink latency of 13.2 seconds. The log did not include the denominator of scheduled transmission opportunities, so the run cannot provide a packet-delivery ratio. The authors limit the hardware conclusion to execution validation, with no claim of end-to-end energy efficiency, field-scale RF behavior or multi-hop operation.
The energy figures in the analysis are modeled primitive costs, not end-to-end measurements. One acknowledgment-free scheduled transfer was modeled at 79.2 microjoules for the sender and 165 microjoules for the receiver; initial alignment added 76.8 microjoules across the pair. A 30-second maximum listening window at 12 microwatts was counted as 360 microjoules. The calculation excluded regulator loss and nonvolatile-memory commits.
MagPie's clearest result is at first contact: under the tested settings, it consistently found and aligned peers. The remaining evidence is more conditional. Contention depends on matching roles to the number of contenders, delivery falls in variable charging scenarios, and the hardware work validates execution without establishing end-to-end efficiency, ambient operation or multi-hop scale.
Paper data and sources
Original title: Rethinking Battery-free Sensing Communication via Wake-up Radios
Authors: Gaosheng Liu, Kasim Sinan Yildirim, Sajal K. Das, Yingxin Song
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text