Speaker
Description
Spiking Neural Networks (SNNs) offer inherent energy efficiency for edge robotics, but their dynamic, event-driven, and sparse nature complicates the provision of hard real-time guarantees. This paper presents a lightweight, custom event-driven scheduler implemented on a resource-constrained Cortex-M0+ microcontroller. The scheduler ensures predictable execution by isolating high-priority spike acquisition from lower-priority cascade processing.
We analyze the Worst-Case Execution Time (WCET) of spike propagation using a sensorimotor neural circuit incorporating a Central Pattern Generator (CPG) driven by a command neuron. This architecture, inspired by rhythmic locomotor networks in flying insects, generates a self-sustained and highly predictable rhythmic output from sparse inputs. Its intrinsic regularity makes it well suited for rigorous evaluation of output stability and temporal jitter within a closed-loop control system.
System performance and timing guarantees are demonstrated by integrating the circuit into a simple mobile robot performing real-time obstacle avoidance. The internal SNN processing load and its WCET are modeled as a self-organized critical (SOC) process using the Bak–Tang–Wiesenfeld (BTW) cascading-event model. By enforcing a bounded refractory period and strong inhibitory synaptic weights, the network operates in a sub-critical regime. This enables derivation of a practical upper bound on maximum avalanche size, thereby establishing deterministic temporal stability of the event-driven cascade under varying sensory loads.
| Minioral | No |
|---|---|
| IEEE Member | Yes |
| Are you a student? | No |