SAT-Edge-Agent: Hardware-in-the-Loop Edge-Agent Orchestration for Onboard Satellite Intelligence
SAT-Edge-Agent is a hardware-in-the-loop edge-agent system deployed on a commercial off-the-shelf ARM-based heterogeneous edge system-on-chip. It uses a browser workspace and FastAPI agent to coordinate a local OpenAI-compatible language service with a YOLO-style oriented-object-detection endpoint returning FAIR1M metadata-backed structured results. Two fixed FAIR1M workloads (single-image and serial two-image) were repeated 20 times each, completing 20/20 attempts. Mean Full-Agent latency was 29.353 s and 60.937 s, with empirical P95 values of 31.166 s and 66.882 s. Mean detector time was 861.386 ms and 1510.920 ms, representing 2.93% and 2.48% of Full-Agent means. Mean CPU utilization was 20.761% and 20.482%. A 200-ms NPU-load field averaged 100% for both workloads, but it represents a shared-accelerator software field rather than detector utilization.
Researchers present SAT-Edge-Agent, a hardware-in-the-loop edge-agent system for onboard satellite intelligence. Deployed on a commercial ARM-based heterogeneous edge SoC, the system integrates a local OpenAI-compatible language service with a YOLO-style oriented-object-detection endpoint. In tests with FAIR1M workloads, the agent completed all 20 attempts for both single-image and serial two-image tasks. Full-agent latency averaged 29.4 s and 60.9 s, with detector time accounting for less than 3% of total latency, indicating most latency occurs outside detector execution. CPU utilization remained around 20%, while a 200-ms NPU-load field averaged 100%, though this reflects a shared-accelerator software metric rather than actual detector load.
The system demonstrates that an edge-agent orchestration layer can reliably translate mission intent into local tool calls on constrained satellite hardware. Profiling reveals that detector execution is a minor contributor to overall latency, suggesting that agent coordination, language model inference, and communication overhead dominate. The use of a shared-accelerator NPU-load field at 100% may indicate polling or software overhead rather than sustained compute utilization, warranting further investigation into true accelerator efficiency.
This work validates the feasibility of deploying autonomous AI agents on edge devices for space applications, potentially reducing reliance on ground-station communication. The low CPU utilization and high NPU-load metric highlight the need for better resource monitoring and optimization in heterogeneous edge systems. The approach could influence satellite mission design by enabling onboard decision-making for tasks like object detection and data triage.
Onboard satellite intelligence can lower operational costs by reducing data downlink requirements and enabling real-time autonomous responses. This technology could be licensed to satellite manufacturers or space agencies seeking to enhance mission autonomy. The use of commercial off-the-shelf hardware suggests a path to cost-effective deployment, though software optimization remains critical for practical use.
Next signals include optimization of the agent orchestration layer to reduce non-detector latency, integration with more diverse sensor payloads, and testing under realistic communication constraints. Further research may explore dynamic workload adaptation and fault tolerance for long-duration missions. Commercial adoption will depend on demonstrating reliability and power efficiency in flight-like environments.