OctoSense: Building a Unified Ecosystem for Open-Source Wireless Sensing

Abstract

As AI enters the physical world, wireless sensing has emerged as a key modality for enabling non-intrusive physical intelligence. However, unlike the mature ecosystems of vision and language models, wireless sensing research remains highly fragmented. The field faces a growing "reproducibility wall" caused by heterogeneous data formats, non-interoperable processing pipelines, and the lack of standardized benchmarks. To dismantle this barrier, we present OctoSense, a unified platform designed to propel wireless sensing toward an "ImageNet-style" research ecosystem. OctoSense introduces a holistic framework that decouples high-level model logic from ad-hoc data nuances through three distinct components: a unified data abstraction for streamlined dataset access, standardized signal operators for efficient model development, and a rigorous benchmark engine for rapid and fair comparison. To support existing datasets and models, OctoSense integrates a comprehensive suite of widely used datasets and models, reducing the effort required for data and model adaptation to a few lines of declarative code. We demonstrate the efficacy of OctoSense through a set of usage examples that reproduce representative models on popular datasets. By providing a foundational open-source infrastructure, OctoSense enables community efforts to accumulate rather than fragment, paving the way for wireless sensing research that is interoperable, reusable, and reproducible by design.