Research paper · 2026

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

OctoSense authors

The University of Hong Kong

A shared interface for wireless datasets, signal processing, and reproducible wireless sensing.

Why is reuse still
so much work?

Wireless sensing has plenty of data.
Working across it is another story.

  • A new dataset. A new loader?

    How can we use multiple datasets without debugging a separate loader for each one?

  • Published code. Will it work here?

    How can we reuse it with a different dataset or setup?

  • My dataset. Where do I start?

    How should I organize signals, metadata, and labels so others can use them?

MeetOctoSense! Figure 1: OctoSense connects dataset selection, signal-processing operations, model execution, and an experiment specification in one workflow.

Semantic-Aware Data Representation

By position

y = STFT(x, axis=0)
STFT along the time axis of real CSIbin
timeframe →
Same spectrum
0time

By semantics (ours)

OctoSense
y = STFT(x, axis="time")
Identical time-axis STFT before and after axis permutationbin
timeframe →
Same spectrum
time0

Initially both methods select time. After the cube rolls, position zero selects subcarrier, while the named operation still selects time.

A dataset is more than a folder of tensors.

A catalog entry describes the dataset. Each sample keeps its own identity, signal context, and target association.

11 datasets

Showing 11 datasets.

Different sources. Reusable processing.

Three Wi-Fi datasets. Shared signal processing. A library of models.

Wi-Fi datasets

Processing Script

process = T.compose([
    T.AxisNormalize(
        "time", method="l2",
        epsilon=1e-8),
    T.Magnitude(),
    AutoAlign(
        model.get_input_contract(),
        time_axis="time",
        flatten_axes=(
            "subc", "tx", "rx"),
    ),
])

Sequence example · each model declares its input contract.

Model library

Make the experiment explicit.

Capture dataset selection, splits, processing, models, and runtime settings in one experiment specification.

operanet-activity.yamlBenchmarkSpec

Explore OctoSense