Sensor-side AI is the find-stage AI in the sensor-to-shooter loop. It runs on UAV payloads, ground vehicles, soldier-worn devices, ground-based radar processors, and tactical-edge servers. The compression it provides — converting raw sensor output into ranked candidate detections in milliseconds — is the most operationally proven AI capability in defense in 2026. Part 2 covers the engineering of this layer: hardware choices, model deployment, the data pipeline, and the failure modes that surface in operations.

The architectural framing is established in Part 1: The Loop and the broader AI discipline in The Complete Guide to AI in Defense Software. This part goes operational.

Why Inference Belongs at the Edge

The argument for edge inference is four-fold and bounded.

Latency. A UAV detection that takes 100 ms at the edge can take 1-5 seconds round-trip to a centralized inference service over a contested link. For tactical scenarios where the loop closes in seconds, this difference is decisive.

Bandwidth. A 4K full-motion video stream from a UAV is megabytes per second. The detection output — a few bounding boxes with classifications and confidences — is bytes per detection. Sending only the inference output back, plus selectively the video chips that triggered them, reduces link load by orders of magnitude.

Resilience. The UAV that loses link continues to detect, classify, and store. When the link returns, the buffered detections flow back. A platform that requires central inference to operate becomes useless during the inevitable link interruptions.

Security. Less raw data leaving the secured device means a smaller attack surface, simpler classification handling, and fewer egress points for the security review. For sensitive sensors operating at high classification levels, edge inference also keeps the model proprietary — the weights stay on the device.

The broader edge-AI use-case landscape is treated in Edge AI Military Use Cases.

Hardware: What Runs Where

The hardware choices for edge-AI in defense are constrained by power, thermal envelope, ITAR considerations, and the physical platform that hosts the inference.

UAV payloads. NVIDIA Jetson Orin family dominates for tactical and operational UAVs. Strong inference throughput per watt, mature TensorRT integration, well-understood by integrators. ITAR positioning matters in European programmes — see ITAR-Free Defence Software. Smaller UAVs increasingly use Qualcomm QCS-class SoCs or dedicated NPUs (Hailo, Ambarella) for better performance-per-watt at small budgets.

Ground vehicles. Higher power and thermal headroom than UAVs allow ruggedized server-class GPUs (NVIDIA L4, RTX A-series) or multi-Jetson clusters. The vehicle's compute envelope is bounded by the alternator and the cooling system rather than by component constraints.

Soldier-worn devices. Strict power and thermal budgets push to dedicated NPUs, often integrated into the device's main SoC (Qualcomm, MediaTek industrial variants). Inference latency is constrained by the small model that fits.

Tactical-edge servers. Ruggedized 1U or 2U chassis at company or battalion command posts. Multiple GPUs, multiple TB of storage, run multiple model instances. The same chassis is the deployment target for fusion, COP backends, and central inference services that the smaller edge devices cannot run.

The detailed hardware comparison and selection criteria are in Edge AI Hardware Comparison.

The Model Pipeline: Train Centrally, Infer at the Edge

The pattern is consistent across edge AI in defense. Train at high precision in a secure data centre with aggregated multi-source data. Quantize and convert to the target inference runtime. Validate against deployment-environment data. Deploy as signed artefacts to the edge platforms.

The stages, in engineering detail:

Training data preparation. Defense training data is scarce, classified, biased toward whatever the available sensors happen to have captured, and unevenly labelled. The investment in labelling, provenance tracking, and class balance is structural — without it, downstream accuracy claims are meaningless. Synthetic data fills gaps; never alone. See Synthetic Data for Defense AI.

Training. Conventional PyTorch or TensorFlow training in a secured environment. Models are vision transformers, YOLO-family detectors, or specialized architectures depending on sensor modality. Hyperparameters and architectural choices are tracked alongside the model weights for reproducibility.

Quantization. FP32 training, INT8 or INT4 deployment. Quantization-aware training where accuracy degradation is unacceptable. The accuracy delta between training precision and deployment precision is measured on the deployment-environment validation set — not on the training set, which would understate the regression.

Conversion. ONNX as the interchange format. TensorRT for NVIDIA targets, vendor-specific runtimes elsewhere (Qualcomm SNPE, Hailo runtime, etc.). The conversion pipeline is automated end-to-end; manual conversion never survives the first model update. The pattern is in ONNX and TensorRT Model Optimization.

Validation. The converted, quantized model is validated against a representative validation set that matches the deployment environment. A model that performs well on lab-curated data and badly on operational data is operationally useless — and almost guaranteed unless the validation set reflects reality.

Deployment. Signed artefacts deployed to edge platforms through the C2 software supply chain. Update cadence matched to the operational rhythm and the platform's maintenance windows. Rollback paths tested, not assumed.

The Edge Data Pipeline

The edge AI does not operate in isolation. It is part of a data pipeline that closes the loop with central training.

Outbound from the edge. Detection events flow back to the C2 fusion engine as candidate tracks. The detection-to-track adapter pattern is the one from Building a C2 System, Part 2: The Fusion Engine.

Selective full-data return. When bandwidth allows, the video chips, audio segments, or signal samples that triggered high-confidence detections flow back centrally. These become future training data. Selectivity is policy-driven — operator confirmation rate, novelty detection, deliberate sampling for hard examples.

Active learning feedback. Where operators correct detection labels — false positives flagged, missed objects added — the corrections feed back to the training data store with provenance. This closes the loop between operations and model improvement.

Drift monitoring. The platform tracks model performance over time. Confidence distributions, classification distributions, and operator-correction rates surface drift before it becomes a deployment problem. Drift detection triggers re-training, not silent degradation.

Sensor Modalities: Different Sensors, Different AI

The same architectural pattern instantiates differently across sensor modalities.

Electro-optical (EO) and infrared (IR) imagery from UAVs. The canonical edge-AI application. Object detection, classification, tracking through video. Engineering treated in Computer Vision in Defense Systems.

Synthetic-aperture radar (SAR). Detection from SAR imagery has matured rapidly. Deep-learning detectors trained on SAR-specific features (speckle patterns, scattering signatures) outperform classical methods at the moving-target indicator (GMTI) and change-detection tasks.

SIGINT and ELINT. Signal classification, modulation recognition, emitter fingerprinting. Deep learning has displaced or augmented classical feature-engineering for many of these tasks. See SIGINT Platform Components.

Acoustic. Gunshot detection, vehicle classification, UAV detection by acoustic signature. Edge-deployable on small power budgets — acoustic ML is computationally lighter than vision ML.

AIS and ADS-B. Anomaly detection on civilian-broadcast tracks. Spoofed AIS detection, behavioural anomaly identification. The pattern is in Integrating AIS and ADS-B into a Military Picture.

Key insight: The model that works in the lab is not the model that survives operations. Lab models often have higher headline accuracy and lower operational accuracy because their training data does not include adversarial inputs, sensor jamming, weather degradation, and the rare-but-critical edge cases. The operational model is a continuously-improved descendant of the lab model — not the lab model itself.

Failure Modes in Operations

Edge AI models fail in operations in patterned ways. Engineering the platform to surface these failures rather than mask them is half the operational discipline.

Distribution shift. Operational sensors see things the training data did not — new vehicle types, new camouflage patterns, novel weather conditions. The model returns confident-but-wrong classifications. Mitigation: drift monitoring, out-of-distribution detection, conservative confidence calibration, fast re-training when drift is detected.

Adversarial inputs. Deliberate manipulation of the sensor input to mislead the classifier. Adversarial patches, sensor spoofing, deepfake imagery. Mitigation: adversarial training, ensemble methods, sanity checks against physics-based priors. The robustness discipline must be built in from sprint one; retrofitting it is unreliable.

Sensor degradation. Jammers, fog, dirt on the lens, intermittent power. The model receives degraded input and produces degraded output, often with miscalibrated confidence. Mitigation: explicit degraded-input detection, confidence calibration against degraded examples in training, graceful-degradation paths to operator review.

Hardware failure. The Jetson overheats. The NPU loses sync. The inference returns nonsense or zero. Mitigation: health monitoring, watchdogs, fallback to lower-fidelity inference, fast failover to other nodes.

Model staleness. The deployed model is six months old; the threat picture has evolved. Mitigation: scheduled re-training and re-deployment, operationally-driven update cadence, the deployment pipeline that gets a new model into the field in days rather than months.

When Federated Learning Helps

Federated learning is the right technique when training data exists across multiple sites that cannot share it. In defense, the canonical case is multi-national coalition data — radar tracks observed by allied nations, sensor inputs from partner forces — that cannot be centralized for legal, classification, or sovereignty reasons.

The pattern: each site trains locally on its own data; only model gradients or weight updates leave the site. A coordinator aggregates updates into a global model. The classified data never moves. The engineering pattern, including secure aggregation and byzantine-robustness considerations, is in Federated Learning for Military Sensors.

Federated learning is operationally complex. Setup time is long; trust between participating sites is foundational; accreditation of the coordinator is non-trivial. The technique deploys when the centralization alternative is impossible or politically blocked. It is not the right tool for single-nation training pipelines, however appealing the architecture sounds in a slide deck.

What's Next

Part 2 has covered the find-stage AI. Sensors produce detections, edge inference filters and ranks them, the data pipeline closes the loop with central training. The platform now has trustworthy candidates flowing back to operators.

Part 3 takes the candidates and goes operational on the target stage. Decision-support AI, recommended-engagement lists, course-of-action analysis, LLM-augmented briefing tools — and the structural patterns that keep humans in the loop while AI compresses cognition.