Wireless sensor and machine-to-machine networks usually share the air in one of two ways. Devices can contend for the channel, listening and transmitting when it seems free (carrier-sense multiple access, CSMA). Or a coordinator can hand out reserved time slots (time-division multiple access, TDMA). Contention is cheap when traffic is light and collapses under load. Reservation is collision-free under load and wasteful when traffic is sparse. Hybrid medium access control (MAC) protocols split each frame into a contention part and a reserved part and move the boundary between them as load changes. The hard part is the rule that moves the boundary. Heuristic thresholds are cheap but oscillate when bursts end. Convex re-optimization is principled but needs device counts and traffic statistics that a small node does not have. We propose to set the boundary with the dynamics of a spiking neuron. Each node is represented at the coordinator by one leaky integrate-and-fire unit: a counter that accumulates evidence, decays over time, "fires" when it crosses a threshold, and then stays silent for a refractory period. Here, firing means "grant this node reserved slots," and the refractory period means "hold the grant."
The second direction, neuromorphic energy efficiency, enters as the accounting lens rather than as a sales pitch. Spiking computation costs energy roughly in proportion to how many events it processes. Hybrid MAC traffic is event-sparse at exactly the loads where radio energy matters most. We will therefore measure, on real nodes, whether the controller's own energy stays below the radio savings it produces across the whole load range. We will also check whether that result depends on neuromorphic silicon at all, or only on the event-driven control law.
Core Claim
The claim has two parts, and it is falsifiable. First, a population of spiking units with leak, threshold and refractory recovery gives a hybrid CSMA/TDMA coordinator three things threshold rules lack. It has built-in hysteresis, which removes the oscillation Queue-MAC shows after bursts. It has an explicit stability condition under the one-frame feedback delay that every beacon-based hybrid MAC incurs. And the reserved period shrinks smoothly to zero under light load. Second, because the controller only does work when channel events occur (packet arrivals, observed collisions, idle contention slots), the combined energy of radio and controller per delivered packet should improve on both threshold and optimization baselines. This holds only if the controller's leakage floor stays small relative to the radio's sleep power. We expect the second part to depend on the platform, and we treat it as something to measure rather than assume.
The closest prior work is Network-Optimised Spiking Neural Network (NOS) Scheduling for 6G O-RAN. NOS models each cellular user as a two-state spiking unit. Its excitability tracks queue backlog, its recovery enforces fairness, and interference arrives as delayed spikes weighted by an interference graph. It proves queue stability when a spectral margin is positive. NOS allocates resource blocks in a cellular grid. It has no contention phase, no hybrid frame and no hardware energy measurement, and it leaves gain adaptation open. We reuse its modeling idea and its style of stability analysis. We move both into the setting where the controlled quantity is the contention/reservation boundary, where the feedback delay is the frame length, and where energy is measured in joules on a node.
Research Program
The controller. The coordinator keeps one unit per associated node. Each unit's potential integrates three inputs. The first is that node's piggybacked queue indicator, which is the same one-byte field Queue-MAC uses, so no header change is needed. The second is a count of collisions and idle slots observed during the contention period, updated per slot as events. The third is inhibitory input from other units, which prevents over-allocation of the reserved period. When a unit fires, its node receives reserved slots in the next frame, sized by the reported backlog. That frame's reserved-period length is the sum of the granted slots. Refractory recovery keeps the grant for a few frames and then releases it, so a finished burst does not leave reserved slots sitting idle. With no firing units, the frame falls back to pure contention.
The stability analysis. The coordinator acts on information one frame old. We will adapt the NOS spectral-margin argument, which there takes the form $\delta = k_\star(\Delta) - g H \rho(W)$, to this setting:
- $\Delta$ becomes the frame duration.
- $g$ is the input gain on queue evidence.
- $H$ is the corresponding system constant.
- $\rho(W)$ is the spectral radius of the matrix coupling the units. In a single-hop cell, this coupling comes from the shared contention period. In a cluster tree, it comes from hidden-terminal adjacency.
The deliverable is a condition, checkable offline, that tells a deployer which gains keep queues stable for a given frame length. It will be verified empirically by sweeping gains across the predicted boundary.
Joint energy accounting. We will define the per-packet energy of a node-plus-coordinator system as
$$E_{\text{pkt}} = \frac{1}{N}\Big(E_{\text{radio}} + P_{\text{leak}}\,T + e_{\text{ev}}\,N_{\text{ev}}\Big),$$
where each term means:
- $N$ is the number of packets delivered in a run of duration $T$.
- $E_{\text{radio}}$ is transceiver energy summed over sleep, listen, receive and transmit states.
- $P_{\text{leak}}$ is the controller's static power.
- $e_{\text{ev}}$ is the energy per processed event.
- $N_{\text{ev}}$ is the number of events processed.
The equation makes the constraint explicit. Energy-efficient neuromorphic classifiers found that leakage was 95% of TrueNorth's consumption at low activity. So the event-proportional term only matters if $P_{\text{leak}}T$ is small against radio sleep energy. The same controller will be implemented in two ways and measured. One is integer interrupt-driven code on the STM32W108 microcontroller used by Queue-MAC. The other is a neuromorphic core, if one is available to us. For scale, the file reports about 17 µW for a small recurrent network on TrueNorth in Conversion of Artificial Recurrent Neural Networks to Spiking Neural Networks, and 140x lower energy per inference than a Jetson TX2 on Loihi in PopSAN. Those figures come from other workloads and serve only as priors. We will not treat them as our results.
Other hybrid MAC work mostly sets the scope. The four-period frame of A scalable Hybrid MAC protocol for massive M2M networks supplies our frame template: notification, contention, announcement and reserved transmission. The multi-group wake-up-radio design of MG-HYMAC shows that radio-side savings can be large (35–64% lifetime gains over HyMAC/CPMAC). Any controller cost therefore has to be charged against a strong radio baseline, not a naive one. We deliberately leave aside the hybrid TDMA-NOMA/RSMA physical-layer schemes and fully beaconless decentralized coordination. Decentralized coordination is a natural follow-on once the centralized stability result holds.
Baselines
The naive reference is fixed-window CSMA with a 20 ms period, together with standard IEEE 802.15.4, as in the Queue-MAC testbed. On that testbed, Queue-MAC reports that 802.15.4 and CSMA drop 40–60% of packets to queue overflow at high node counts. It also reports that Queue-MAC cuts maximum packet delay by up to 14.7x relative to the fixed 20 ms CSMA protocol. That 14.7x maximum-delay reduction is the number to beat. Queue-MAC reports no numeric energy per packet, so our energy comparison will be measured fresh on identical hardware.
Queue-MAC itself is the strongest practical competitor. It uses the same one-byte information, runs on the same microcontroller and needs no training. We will run it with its published threshold rule and with a threshold grid-searched on our traffic under the same tuning budget as our gains. Beating it means lower maximum and P99.9 delay, and less delay oscillation after bursts, at equal or lower measured energy per packet.
The principled competitor is the convex frame optimization of A scalable Hybrid MAC protocol for massive M2M networks. It chooses the optimal contention probability and admitted-device count per frame. We run two versions. One is an oracle version given the true active-device count. The other is a realistic version estimating the count from the previous frame, with the cost of re-solving charged to the energy budget. No reported numeric throughput or delay result was found on Lacuna for this method. Matching the oracle under stationary Poisson traffic and beating the realistic version under bursty traffic would count as success.
For a learned competitor, we will train a small recurrent Q-network policy in the style of CS-DLMA, outputting the boundary per frame from the same inputs. It will be quantized to run on the microcontroller. Its energy per decision is charged in full.
Two ablations isolate the key ingredient. Removing refractory recovery and leak reduces the controller to a plain threshold on accumulated evidence. The interrupt-driven integer implementation of the identical dynamics separates "the control law helps" from "neuromorphic silicon helps."
Why not just…?
Why not just tune Queue-MAC's thresholds? Tuned thresholds may match us on stationary Poisson traffic, and we expect that. Queue-MAC's own authors, however, note fixed-packet-length assumptions and latency oscillation when bursts end. A threshold without memory or hysteresis cannot tell a burst's tail from a new burst. The refractory ablation tests exactly this. If the no-refractory variant performs as well as the full controller under compound-Poisson bursts, the spiking dynamics add nothing and we will say so.
Why not just re-solve the convex frame optimization each frame? It is optimal only when the device count and traffic statistics are known and stationary. On a node it means solving a problem per frame with estimated inputs. The oracle-versus-realistic comparison measures how much of its advantage survives estimation error under bursts. The energy accounting measures what the re-solves cost. If the realistic optimizer matches our delay tails at lower total energy, our first claim fails.
Why not just run an exponential moving average and a threshold on the microcontroller, with no neuromorphic hardware? This is the most serious threat. A leaky integrator is, algebraically, an exponential moving average updated on events. We therefore make this the central ablation rather than argue around it. Our prediction is that the control law, not the substrate, delivers the delay and stability gains. We also predict that the neuromorphic core pays off only where its leakage floor falls below microcontroller idle power, a regime Energy-efficient neuromorphic classifiers suggests may be narrow. If the microcontroller version wins everywhere, the paper's energy conclusion becomes "event-driven control, implemented conventionally," and that is a useful negative result for the neuromorphic community.
Why not just use deep reinforcement learning with decisions at deferral events, as FC-MAPPO does? AI-Enhanced Distributed Channel Access for Wi-Fi 8 reports collisions reduced from about 40% to about 34% and a 10% throughput gain with 10 dense stations. It achieves this with a dense transformer that must coarsen decisions to millisecond-scale virtual slots, and it addresses pure contention only. Our quantized recurrent-policy baseline represents this family at a size that fits a sensor node. Our experiment shows whether it matches our delay tails, and whether its per-decision energy erases its gains.
Proposed Experiments
The primary testbed reproduces the Queue-MAC hardware setup: STM32W108 nodes, Poisson arrivals with 500 ms mean interval, 95- and 120-byte payloads, and 1 to 40 nodes. We extend it with compound-Poisson bursty traffic of the kind used in the NOS evaluation and with mixed payload sizes, since fixed payloads are a stated weakness of Queue-MAC. For multi-hop coupling, we build interference graphs mirroring the NOS pair2, line4 and ring8 topologies as cluster trees with hidden terminals, and run matching simulations to 200 nodes for M2M scale.
We will report the following metrics:
- maximum and P99.9 packet delay;
- delivery ratio and queue-overflow loss;
- a burst-end oscillation measure: variance of the reserved-period length in the frames after a burst ends;
- measured energy per delivered packet, split into radio, controller static and controller event terms.
Three checks decide the proposal. First, a gain sweep across the adapted spectral-margin boundary should show queues staying stable on the predicted side and diverging on the other. If the boundary does not predict behavior, the theory contribution fails. Second, the refractory-removed ablation should show burst-end oscillation and fatter delay tails that the full controller does not. If it does not, the dynamics are unnecessary. Third, the two-implementation energy comparison, plotted across offered load, will locate the break-even point where controller energy stops being negligible. It will state plainly whether any load regime exists in which the neuromorphic substrate beats the event-driven microcontroller implementation.
References
- Network-Optimised Spiking Neural Network (NOS) Scheduling for 6G O-RAN: Spectral Margin and Delay-Tail Control
- Queue-MAC: A queue-length aware hybrid CSMA/TDMA MAC protocol for providing dynamic adaptation to traffic and duty-cycle variation in wireless sensor networks
- Energy-Efficient Neuromorphic Classifiers
- Conversion of artificial recurrent neural networks to spiking neural networks for low-power neuromorphic hardware
- Deep Reinforcement Learning with Population-Coded Spiking Neural Network for Continuous Control
- A scalable Hybrid MAC protocol for massive M2M networks
- Energy-Aware Hybrid MAC Protocol for IoT Enabled WBAN Systems
- Carrier-Sense Multiple Access for Heterogeneous Wireless Networks Using Deep Reinforcement Learning
- AI-Enhanced Distributed Channel Access for Collision Avoidance in Future Wi-Fi 8