邊緣 AI 與航太
SAT-Edge-Agent Tests a Satellite Edge Agent, with Object Detection Accounting for Only About 2.5% of End-to-End Latency
SAT-Edge-Agent integrates a local language service and an oriented object detector on a commercial heterogeneous ARM SoC, completing all 40 runs of a fixed workload. Measurements show that most latency comes not from the vision model but from agent orchestration and the language service, although the study did not evaluate detection accuracy, power consumption, or flight reliability.
The constraints on satellite-based AI extend beyond model size: task parsing, tool calls, streaming responses, and state persistence can also dominate real-world latency. SAT-Edge-Agent provides a hardware-in-the-loop prototype deployed on a commercial heterogeneous ARM edge SoC. Using a browser workspace and a FastAPI agent, the system orchestrates a local OpenAI API-compatible language service before invoking the project’s internal YOLO-like oriented object detection endpoint. Detection results are returned in a structured format together with metadata from the FAIR1M remote-sensing image dataset.
The study evaluated two fixed workloads—one using a single image and another using two consecutive images—with 20 repetitions each. All 40 runs completed successfully. For the single-image workload, mean end-to-end agent latency was 29.353 seconds, with an empirical P95 of 31.166 seconds. For the two-image workload, the mean was 60.937 seconds and the P95 was 66.882 seconds. By comparison, the detector took an average of only 861.386 milliseconds and 1.511 seconds, respectively, accounting for 2.93% and 2.48% of the complete workflow. This indicates that measuring only the vision core running on the NPU would substantially underestimate delays caused by task-level processing, language inference, and data orchestration.
Mean CPU utilization was approximately 20.8% for both workloads. The system-recorded NPU field, sampled every 200 milliseconds, averaged 100%. However, the authors explicitly warn that this is a software-reported metric for a shared accelerator and must not be interpreted as exclusive detector utilization or as a calibrated measure of hardware utilization. The public evidence package includes de-identified per-request logs, redacted JSON, normalized Server-Sent Events examples, and scripts for recomputing the statistics, with an emphasis on establishing an observable hardware-in-the-loop boundary.
For edge-agent engineering, the results suggest a practical priority: before further compressing the detection model, first break down latency from the language service, orchestrator, serialization, and tool round trips. However, the work tested only two fixed image workflows and did not report detection accuracy, calibrated power consumption, resilience in radiation environments, recovery from connectivity loss, or flight certification. It demonstrates that the prototype can execute repeatedly, but it does not establish that the system is ready for autonomous decision-making in orbit.