Normalize heterogeneous warehouse sensor events into a single sparse state-update stream, then let HST schedule only the affected computation. Once events are normalized, HST doesn't care whether a change came from a camera, an RFID antenna, or a barcode scanner.
A warehouse has M state cells; analytics operator
A (M→N) maps state to derived quantities. Each tick a sparse batch of
events perturbs cells by dx; we recompute A·(x+dx) as a
full mat-vec vs HST scheduled delta.
| event mix | dirty cells/tick | dirty tiles/tick | full recompute | HST scheduled | speedup |
|---|---|---|---|---|---|
| rfid_sparse | 14.3 | 2.0 | 556.7 ms | 56.2 ms | 9.91× |
| barcode_station | 21.3 | 3.0 | 558.6 ms | 81.9 ms | 6.82× |
| asset_location | 28.4 | 4.0 | 558.1 ms | 97.9 ms | 5.70× |
| mixed_warehouse | 42.5 | 6.0 | 561.1 ms | 141.2 ms | 3.97× |
| camera_light | 87.6 | 16.0 | 570.6 ms | 328.9 ms | 1.73× |
| camera_heavy | 240.9 | 62.9 | 576.9 ms | 1245.9 ms | 0.46× |
Answers "hundreds of barcode readers + RF antennas vs cameras":
→ RFID / barcode / asset location: 5–10× — sparse, tile-local, HST's home turf.
→ blended + light camera analytics: ~2–4×.
→ dense per-frame camera object churn: <1× — the dirty set stops being sparse; full recompute wins.
Mirrors the measured RTSP-CV finding. Every path validated rel_err ≈ 1e-16 against full recompute.