Radio environment maps¶

isaac_net.tools.rem samples the large-scale radio of a configuration on an x–y grid and writes a radio environment map (REM), like the REM helper of 5G-LENA. It uses radio.RadioMC, the same radio the NR engine and NetSlotMC use, so every channel model works (log_distance, tr38901, radio_map) with the gNB layout of NRConfig.gnb_xy().

python -m isaac_net.tools.rem --out rem.npz --png rem.png --res 2 --preset multicell --set n_cells=3 channel=tr38901_umi
isaac-net-rem --out rem.npz --bounds 0 0 200 100 --res 1 --set channel=radio_map radio_map_path=map.npz n_cells=2 cell_positions_m="((25,75),(125,75))"

from isaac_net.core import multicell
from isaac_net.tools.rem import compute_rem, load_rem, save_rem

rem = compute_rem(multicell(3, channel="tr38901_umi"), resolution_m=2.0, seed=0)   # dict of numpy arrays
save_rem(rem, "rem.npz", png="rem.png")                                            # png needs matplotlib
rem = load_rem("rem.npz")

Array	Shape	Meaning
`x`, `y`	`[W]`, `[H]`	grid cell centres (m, env-local); rows run along y
`gnb_xy`	`[C, 2]`	gNB positions
`pathgain_db`	`[C, H, W]`	large-scale gain of every gNB-point link: path loss, shadowing and, for `tr38901`, the LOS state (negative dB)
`rsrp_dbm`	`[C, H, W]`	DL RSRP per resource element: `gnb_tx_dbm` over the 12 · `dl_nprb` REs of the DL carrier, plus the path gain
`sinr_db`	`[H, W]`	best-cell DL SINR per PRB with every gNB transmitting on every PRB (full-buffer interference, as 5G-LENA's REM), over the UE noise floor `noise_dbm_per_prb("ue")`
`serving`	`[H, W]`	serving cell = argmax RSRP, the engine's max-RSRP attach
`los`	`[C, H, W]`	LOS state, only for channel models that have one (`tr38901`)
`meta`	string	JSON: config summary, bounds, resolution, seed, env, noise floor, DL PRBs

Every grid point is a receiver at ue_height_m. The shadowing and LOS fields are those of env env of a RadioMC whose fields come from a torch.Generator seeded with seed, so a fixed seed gives the same map on every run (the tool runs on the CPU). The fields are those of the channel model, not of a particular engine run: an engine with the same config draws its own fields from its engine RNG. The map is outdoor coverage without robots: the blockage add-on (other robots as spheres) and O2I (an indoor draw per receiver) are turned off, since grid points are not robots.

Options of the command line: --preset names a preset of isaac_net.core.config (multicell, lena_like, ...), --set FIELD=VALUE ... overrides NRConfig fields (values are Python literals), --bounds X0 Y0 X1 Y1 (default: the arena (0, 0, cell_arena_m, cell_arena_m)), --res the grid spacing in m, --seed, --env, --png (an optional figure, needs matplotlib) and --panels (the comma-separated panels of that figure among rsrp, sinr, serving, los and pathgain, default rsrp,sinr,serving,los). The figure is drawn by isaac_net.viz.maps.plot_rem (Plots and reports), with the gNBs marked and a scale bar, and the default panel set drops los without a warning when the channel has no LOS state. From Python, shape=(H, W) fixes the grid size instead of the spacing, and radio_map= passes a RadioMap directly.

Tests: tests/test_dl_traffic_fdd_rem.py (grid shape, serving cell = argmax RSRP, determinism for a fixed seed, a different map for another seed, the npz round trip, the command line).