Research: camera-free electromagnetic world model via privileged multimodal distillation

Goal

Move RuView beyond activity classification toward a camera-free electromagnetic world model that predicts geometry, motion, and channel state from RF observations at deployment time.

Key constraint

Do not require cameras in production. RGB, LiDAR, depth, UWB, or FMCW may be used as privileged training evidence only.

Proposed pipeline

CSI + LiDAR/depth/RGB during training -> shared electromagnetic representation -> teacher -> distillation -> CSI-only student -> geometry + motion + channel

Workstreams

1. Define synchronized multimodal capture schema with timestamp, pose, RF metadata, calibration, and environment ID.
2. Build teacher representation that jointly predicts RF reconstruction and spatial geometry.
3. Distill into CSI-only student with uncertainty outputs.
4. Test cross-room and cross-hardware generalization.
5. Feed spatial hypotheses into RuVector/RuField/WorldGraph rather than treating outputs as isolated classifications.
6. Compare WiFi CSI against UWB/FMCW where hardware is available.

Metrics

• depth/geometry error
• motion tracking error
• CSI reconstruction similarity
• cross-room degradation
• cross-device degradation
• uncertainty calibration
• inference latency on edge targets

Safety/product constraint

Camera-derived supervision must be disposable after training. Production inference must be demonstrably camera-free for deployments making that claim.

Acceptance criteria

A CSI-only student reconstructs a stable coarse spatial representation in held-out rooms and on at least two RF hardware families, with measured uncertainty and no runtime dependency on visual sensors.

The issue names no files, tests, or entry points, so begin by locating the existing CSI ingestion, training, and spatial-output components. Use the proposed synchronized multimodal schema and CSI-only deployment constraint as the initial scope; done requires held-out-room and two-hardware evaluation with uncertainty, latency, and no runtime visual-sensor dependency.