Systems engineering
The SWARMind drone OS and algorithm stack for technical evaluators: distributed cognition, Tiny AI compression, decision-level fusion, Byzantine consensus, self-defence, and graceful degradation — as specified in the proposal corpus.
SWARMind treats each airframe as a full software peer: it perceives, runs on-board algorithms, and contributes to swarm-wide consensus without a central aggregator. That is how a swarm operating system stays resilient (no central failure point), privacy-preserving (raw sensors stay on-drone), and bandwidth-efficient (decision vectors, not raw streams).
Baseline algorithms: YOLOv8n for object detection and MobileNetV3-Small for scene classification. INT8 post-training quantisation yields ~4× size reduction and 2–4× latency improvement on INT8 hardware, with theoretical mAP cost of 0.5–2%. Hardware-aware NAS (MCUNet-style) targets Pareto-optimal accuracy–latency–memory trade-offs. Software KPIs: compressed model <20 MB; inference <100 ms on Jetson Orin NX / Zynq stack; detection >85% mAP.
Each drone runs the full local inference pipeline and broadcasts a decision vector ≤512 bytes/s encoding detections, environmental scalars, node health, and consensus sequence. Primary physical-swarm link: 5.8 GHz OFDM (up to 10 Mbps). 868 MHz LoRa is the long-range redundancy fallback under ETSI duty-cycle constraints. End-to-end decision latency target: <200 ms.
Gossip epidemic broadcast converges logarithmically with swarm size. Lightweight PBFT-adapted quorum: for N=20, f≤6 Byzantine faults (N≥3f+1); detection confirmed at 2f+1=13 consistent peers. Resilience specification: mission integrity under ≥40% packet loss and ≥30% simultaneous node failure — validated in ROS 2/Gazebo with fault injection before field pilots.
On-device hostile recognition combines lightweight vision, RF pattern analysis, and behavioural prediction. Countered include cooperative evasion (threat probability <60%), active blocking / swarm jamming (≥60%), and autonomous fallback when links degrade — non-lethal cooperative response as framed in the concept papers.
Three-layer OS partitions: if the AI processor fails, the FPGA continues EKF fusion; if both upper layers fail, the STM32H7 MCU executes a pre-loaded Return-to-Home from non-volatile memory. Cognitive function is distributed in software and hardware — not concentrated in one binary.
Reference hardware for the software stack (proposal): Jetson Orin NX 16 GB · Xilinx Zynq UltraScale+ ZU3EG · STM32H7 · RGB/thermal/LiDAR/IMU/RF/Galileo-EGNOS · custom hexacopter ~850 g payload · >20 min endurance KPI.
Systems audio
Play a short boot chime and low avionics hum while you browse.
