Research: CSI optical flow as a dense RF motion field primitive

Finding

WiFlow: Estimating Optical Flow using WiFi Channel State Information, arXiv:2609.02452, announced 2026-09-03, introduces a CSI optical-flow dataset and three architectures intended to estimate dense motion fields from WiFi CSI rather than camera frames. The project reports code and data at https://visinf.github.io/wiflow.

Claim class: originating-team newly released. No RuV performance result is claimed here.

Why it matters

RuView currently has multiple sensing heads above CSI. A dense motion field could be a more reusable intermediate representation than another activity classifier because motion can feed tracking, pose, occupancy, fall/mobility analysis, world-state updates, sensor fusion, and camera-free spatial inference.

RuV mapping

RuView ingestion and calibration -> CSI encoder -> RF motion field -> RuField/WorldGraph temporal update -> RuVector memory -> downstream task heads.

Potential consumers include pose/activity heads, WorldGraph motion hypotheses, LiDAR fusion, UWB/FMCW fusion, and edge visualization.

Reproduction first

Pin the WiFlow repository and dataset version. Record CSI hardware, antenna geometry, carrier/bandwidth, packet rate, synchronization, room layouts, labels, train/test split, model parameters, software versions, accelerator, and license before implementation.

Compare:

1. existing RuView motion/activity representation
2. a simple temporal CSI baseline
3. WiFlow reference architecture
4. a reduced edge candidate only if the reference reproduction succeeds

Metrics

Optical-flow endpoint error or the paper's released primary metric, cross-room degradation, cross-person degradation, cross-hardware degradation where possible, temporal stability, downstream tracking/pose gain, calibration, p50/p95 latency, memory, model size, power, packet loss sensitivity, missing subcarrier sensitivity, and total sensor cost.

Falsification

The optical-flow representation should be rejected as a core primitive if it improves its own pixel/field metric but does not improve at least two downstream RuView tasks on held-out rooms, or if a simpler temporal embedding performs within variance at materially lower edge cost.

Promotion

Require a reproduced reference result plus at least 10% relative gain on two held-out downstream motion/spatial metrics or equivalent quality with at least 30% lower end-to-end inference cost. Camera data may be used for labels during research but must not become an undeclared production dependency for camera-free claims.

Keep this isolated and reversible. No production sensing claim until real RuView hardware reproduces transfer.

Start by pinning the WiFlow repository and dataset versions, then record the hardware, data, split, model, software, accelerator, and license details listed in the issue. Reproduce the reference result and compare it with RuView's existing motion representation and a simple temporal CSI baseline. Done means meeting the stated downstream-task, latency, cost, and held-out-room promotion criteria without making camera data a production dependency.