Edge AI Deployment Beyond Models: A BSP-Aware Systems Framework for Industrial Embedded Platforms

This academic paper presents a systematic framework for deploying AI models on industrial embedded systems, arguing that successful Edge AI requires treating deployment as a holistic systems problem rather than a late-stage packaging task. The five-layer framework addresses hardware, BSP/OS adaptation, runtime acceleration, application inference, and operations/validation, with implications for reproducibility and field reliability in long-lifecycle industrial products.

The paper identifies a critical gap in how industrial organizations approach Edge AI deployment. Traditional development sequences prioritize model demonstration over platform constraints, creating friction when targeting embedded systems with vendor-specific kernels, heterogeneous accelerators, and long product lifecycles. This disconnect between model-first development and platform-first constraints produces reproducibility issues, unreliable field performance, and difficult diagnosability in production environments.

The proposed BSP-aware framework acknowledges that a deployed AI model operates within a complex execution chain spanning from sensor input through board support package layers to production service loops. By grounding analysis in vendor documentation from Android, NXP i.MX, NVIDIA Jetson, ONNX Runtime, and TensorRT, the authors provide practical architecture patterns that address real deployment scenarios. The framework's emphasis on sustained throughput and field reliability reflects lessons learned from managing heterogeneous edge fleets at scale.

For organizations developing embedded AI applications, this framework reduces deployment risk by forcing earlier consideration of platform-level constraints. Industrial manufacturers benefit from improved reproducibility across production units and enhanced diagnosability when field issues arise. The five-layer abstraction enables clearer communication between hardware engineers, systems teams, and ML practitioners who often work in silos. This structured approach particularly matters for safety-critical applications where device instability poses operational risk.

The paper's systematic treatment validates growing industry recognition that Edge AI success requires systems thinking. As embedded AI applications proliferate across manufacturing, automotive, and infrastructure sectors, frameworks addressing integration complexity rather than model optimization alone become increasingly valuable.

• →Edge AI deployment must prioritize systems architecture and platform constraints over model-first development sequences.
• →A five-layer framework connecting hardware through operations enables reproducibility, diagnosability, and field reliability in industrial embedded systems.
• →Vendor-specific kernels, heterogeneous accelerators, and safety requirements make embedded AI a systems problem requiring cross-functional coordination.
• →Structured integration patterns reduce deployment risk and improve sustained throughput in long-lifecycle industrial products.
• →Platform-aware deployment frameworks address critical gaps between academic model development and production field reliability.