Adaptive Precision Inference for Battery-Powered Field Robots in Infrastructure-Poor Environments

Abstract

Battery-powered robots deployed in infrastructure-poor environments face a fundamental tradeoff between inference quality and operational endurance. Static quantization schemes, which fix precision at deployment time, cannot adapt to the varying computational demands of different task phases. We present an adaptive precision inference framework that dynamically adjusts model quantization level based on real-time battery state, task criticality, and environmental complexity. The framework monitors three signals: remaining battery charge, a learned task-phase classifier that distinguishes high-criticality phases (e.g., extraction near a pipe junction) from low-criticality phases (e.g., straight-line navigation), and a visual complexity estimator. During low-criticality phases, the system aggressively quantizes to INT4, while escalating to FP16 during critical manipulation sequences. Evaluated on the SafAI sewer inspection platform, adaptive precision extends operational time by 2.1x compared to fixed FP16 inference while maintaining 96% of full-precision task success rate. We release the framework as an open-source library compatible with ONNX Runtime and TensorRT.

Keywords