Answers are grounded in this paper and cite their sources.
Visual navigation has become a fundamental capability for embodied intelligent agents, but deploying such systems on resource-constrained embedded edge devices remains difficult due to the substantial computation and communication overhead of high-level decision models. This article proposes the Adaptive Model Partitioning and Deployment System (AMSDS), an edge–cloud collaborative framework for reducing end-to-end inference latency in target-driven visual navigation. AMSDS adaptively determines the collaborative inference mode according to the topology of the high-level decision model, model complexity, workload state, and runtime network condition. For models with acceptable partitioning overhead, AMSDS applies offline DAG simplification and efficient cut inference to obtain low-latency partition strategies. For models with overly complex models, or heavy-workload states, AMSDS switches to a semantic-compression path based on a Semantic Map Convolutional Autoencoder (SM-CAE), which reduces transmission overhead while preserving navigation performance. Experimental results show that AMSDS consistently outperforms baseline methods in inference latency, latency speedup, and throughput. Moreover, evaluations under dynamic bandwidth/load traces and time-varying task-arrival conditions demonstrate that AMSDS remains effective under realistic runtime variations.
No takes yet. Share an insight, caveat, or question.
Hao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: