Edge AI Deployment

Problem

Moving trained aerial perception models onto onboard UAV compute with export, quantization and runtime choices that fit power and latency budgets.

Topic reviewed 2026-08-16 · Membership is an editorial index, not a compatibility or performance claim.

Engineering questions

• Which inference runtime matches the target hardware and model format?
• How does the export path (ONNX, TensorRT, OpenVINO, RKNN) preserve model behaviour?
• What quantization and precision trade-offs apply to the deployment target?
• How is deployability verified on the actual device before integration?

Topic map

Stack layers: edge_ai · perception_localization · hardware

Engineering scope and evidence boundary

This topic groups the release-backed edge inference records that OpenFly can verify from their official repository or documentation. The page is an entry point for comparing an inference runtime, an NPU toolkit or an export helper without treating them as interchangeable. Deployment support, supported hardware and performance figures are recorded only when the official source states them; measured throughput, certification and compatibility claims remain outside the page unless a SourceRef records them. Resource cards expose verification status, maintenance signal, source tier and retrieval date so the reader can move from the topic explanation to the underlying evidence.

Topic membership is an editorial index, not a compatibility or performance claim. Verify each resource through its SourceRef, retrieval date, and verification status.