CARI 2026 · paper #65 · deck v2 for quick review

Name, co-author, one line on scope: delivery measurements in a three-floor concrete building, two variables: radio parameters and channel access. (30 s)

Wiring a building is costly; Wi-Fi needs access points on every floor because concrete eats 2.4 and 5 GHz. LoRa at 868 MHz gets through, so one gateway can cover the building. The price is heterogeneity: one node in line of sight, others behind a slab. Today's question is whether that heterogeneity decides delivery, or something else does. (75 s)

Fig. 1 (right): images/setup/transverse_diagram.pdf

Three lines of work. Signal studies document RSSI and SNR well but treat delivery as secondary. Parameter studies use single nodes. MAC studies show scheduling removes collisions and that CAD is unreliable through obstacles, but in controlled setups. Nobody has put contention and multi-floor link disparity in the same building. (75 s)

Section 2 (related_works.tex)

These two questions organise the rest of the talk: slides 7–9 answer RQ1, slides 10–14 answer RQ2. (50 s)

Gateway on Floor 1. Node 20 shares its corridor, in line of sight. Node 10 is one floor down and Node 30 one floor up, each behind a slab and walls. Same hardware everywhere, power and coding rate fixed. (80 s)

Fig. 1 (left): images/setup/nodes_annotated.pdf · Table 1

C1 isolates the radio: one link, no contention. C2 adds sharing without collisions. C3 adds contention. Between C2 and C3 only the access scheme changes. SF9/BW125 was picked for C2 and C3 because it had the steadiest SNR in C1 (backup B3). (75 s)

Left column RSSI, right column SNR, one row per floor, colour by SF. RSSI ranges from −40.2 to −89.7 dBm, close to 50 dB, consistent with 10–20 dB per concrete floor. Yet SNR barely moves: 6.7 to 6.9 dB on average. No link is margin-limited, which already hints at what PDR will do. BW500 adds 1–3 dB of RSSI and costs 1–2 dB of SNR. (70 s)

Fig. 2: images/Characterization/Floor{0,1,2}/Floor*_Static_{RSSI,SNR}_over_Time_Shaded_SF_Separated.pdf

Boxplots by SF on the left, every configuration on the right. Everything is pressed against 100%. SF medians are at or above 99.6% on every floor. The one outlier, 95.7%, is Floor 2 at SF9/BW500, where the BW500 noise penalty meets an NLOS path. All other configs stay above 97.9%. (75 s)

Fig. 3: images/Characterization/Floor{0,1,2}/Floor*_PDR_by_SF_Boxplot_SequenceBased.pdf, Floor*_PDR_by_Config_SequenceBased_Shaded_SF_Separated.pdf

Same message, same delivery, 219 times the airtime. Within SF12 alone, BW and payload give a tenfold range, 330 to 3,285 ms. Under the 1% duty cycle that is the difference between a packet every 1.5 s and one every five and a half minutes, and an SF12 packet costs 219 times the battery. (80 s)

Fig. 4: images/Characterization/ToA_by_Config_Shaded_SF_Separated.pdf

Before comparing delivery: the channel itself did not change between the two four-day campaigns. Same three bands, same separation. So whatever difference we see in PDR next comes from access, not from the environment. (60 s)

Fig. 5: images/Robin/Robin_RSSI_over_Time_By_Node.pdf, images/ALOHA/Aloha_RSSI_over_Time_By_Node.pdf

Flat at 100% for four days. The two NLOS nodes, about 48 dB weaker, are indistinguishable from the LOS node. Remove collisions and link quality stops predicting delivery. The three dips at one moment are missed scheduler triggers: they hit nodes 60 dB apart with no RSSI or SNR change. (70 s)

Fig. 7: images/Robin/Robin_Packet_Delivery_Rate_over_Time_Scatter.pdf

Same nodes, same building, same SF9/BW125. Now the cloud drops and spreads for all three nodes. Put this next to the previous slide: the only change is how nodes take the channel. (70 s)

Fig. 8: images/ALOHA/Node_Block_PDR_over_Time_Aloha_Scatter.pdf

Rows sorted by signal strength. Under scheduling, all three are within a point of each other. Under ALOHA the order flips: the line-of-sight node, 33 dB stronger than Node 10, loses almost 30% of its packets and does worse than both NLOS nodes. Its blocks swing from 5 to 95%: when the channel is clear it delivers, when transmissions overlap it loses. (85 s)

Section 4.2, Uncoordinated ALOHA Performance

Left is what the data shows: with everything else held constant, the access scheme alone costs 15 points of median PDR. Right is why a strong signal does not protect you: nodes behind slabs may not hear each other, so CAD gives little protection, and the capture effect needs tight timing. We did not isolate why Node 20 is hit hardest; that needs per-collision logging. (75 s)

Sections 2, 4.2 and 5

Two answers. Radio tuning matters for airtime and energy, not for delivery, once the link clears threshold. Access matters for delivery, and it can invert the order that link quality would predict. (70 s)

Short-interval periodic traffic was chosen to expose MAC weaknesses; sparse traffic would collide less. Heavy floors may be part of the story: in lighter buildings nodes could hear each other and CAD would work better. (55 s)

Thank the chair and audience. Leave this slide up: the headline numbers and backup links stay visible during questions. (15 s)

Table 2 (experiment_setup.tex)

Fig. 6: images/Robin/Robin_SNR_over_Time_By_Node_Scatter.pdf, images/ALOHA/Aloha_SNR_over_Time_By_Node_Scatter.pdf

Section 4.1, last paragraph of "PDR Across Configurations". Paper check: it says SF12 is "two to four times" as variable as SF9, but on Floor 2 it is 0.46 vs 0.43 dB (about 1.1×). The ratio holds only on Floors 0 and 1, so the slide does not repeat it. Expect this if someone reads the paper closely.

If asked whether the comparison mixes access scheme and load: yes, load differs, because that is how each scheme behaves over the same four days. But all three nodes face the same load in C3, and the LOS node still comes last.

Section 4.2 (packet and block counts)