Environment reference¶
Warning
Pre-implementation. This page describes proposed contracts. Class signatures, parameter types, schema fields, and behavior are subject to change before code lands. Once implementation exists, content here will be regenerated from docstrings or sourced from running tests.
The environment catalog and the per-task OOD (out-of-distribution) axes. An Environment member names a channel propagation setting; an OODAxis member names the transfer dimension a task is scored on. Generalization is the score, so the environment and axis structure is the benchmark. See Generalization for the protocol and OOD protocol for the evaluation mechanics.
Environment members¶
Member |
Description |
rfgen channel model |
Used for |
|---|---|---|---|
|
Dense urban multipath, sub-6 GHz. Rich scattering, many reflected paths. |
Sionna UMa (urban macro) ray tracer. |
Train plus one OOD target for synthetic tasks. |
|
Sparse rural, near-LoS (near-line-of-sight), sub-6 GHz. Fewer reflectors, stronger direct path. |
Sionna RMa (rural macro) ray tracer, or a TDL (tapped delay line) near-LoS model. |
Train plus one OOD target for synthetic tasks. |
|
Real captures from public RF (radio-frequency) testbeds. Not synthetic. |
None; ingested via RealCaptureAdapter. |
Sim-to-real gap column only; never a scored OOD target. |
Note
v0.1 ships with at least two synthetic environments (E_URBAN, E_RURAL) so leave-one-environment-out is live from day one. With only two environments the environment-level statistic has low power; per-environment scores are reported alongside every aggregate, and the v0.1 aggregate is a comparative ranking under a fixed protocol, not a population estimate. See Tasks validation for the small-N caveat.
E_REAL_CAP is fidelity only: real-capture subsets (Colosseum, POWDER, COSMOS) feed the sim-to-real gap column reported alongside synthetic scores, never the aggregate OOD-avg. See Sim-to-real gap.
Per-task OOD axes¶
Each task carries a controlled OODAxis member. That transfer score, not in-distribution accuracy, is the contract. The closed set of axes:
Member |
Transfer dimension |
Example task |
|---|---|---|
|
Unseen channel environment. |
|
|
Unseen device or unit. |
|
|
Unseen frequency band. |
|
|
Unseen signal-to-noise ratio regime. |
|
|
Unseen drone model. |
|
|
Unseen waveform family. |
|
|
Unseen emitter mix or scene density. |
|
|
Unseen scene type or scenario. |
|
Several tasks layer a secondary axis (for example E-LOC-LOS layers FREQUENCY_BAND on ENVIRONMENT). The full per-task wiring, including the readout head and metric for each axis, lives in Task reference. The evaluation mechanics for each axis live in OOD protocol.
References¶
NVIDIA Sionna: Hoydis et al., “Sionna: An Open-Source Library for Next-Generation Physical Layer Research,” 2023, arXiv:2203.11854. The UMa and RMa ray-tracing models behind
E_URBANandE_RURAL.Koh et al., “WILDS: A Benchmark of in-the-Wild Distribution Shifts,” ICML 2021, arXiv:2012.07421. The leave-one-X-out framing the axes follow.
See Also¶
Task reference: the OOD axis wired to each task.
OOD protocol: the evaluation mechanics for each axis.
Generalization: why transfer, not accuracy, is the score.
Datasets: the recipes that realize each environment.