Claim definition
We test specific properties such as timing, bandwidth, channel response, and scene composition separately. A result supports only the property that was measured.
Skip to ATLAS content01 02 03
Science
Validation evidence
Methods, selected results, negative results, and the limits of the claims supported by each dataset definition.
How to read an ATLAS validation claim
Ask three separate questions: do generated signals have the intended properties, do multi-emitter records and labels match the scene definition, and do models trained on this data work on recorded RF? Evidence for one question does not answer the others.
Negative results
Communications scenes were much sparser than the planned sparse/busy mixture. A Counter-UAS model trained on synthetic signals failed to recognize drones in real captures. We publish both failures so researchers can see which assumptions and uses are unsupported.
Simulation boundary
Matching a published specification or separating synthetic classes does not prove that a model will work on recorded signals. That requires validation with over-the-air data.
Practical implication. Use the dataset when the measured signal and scene properties cover the conditions your task depends on. If success depends on field performance, test with your own representative over-the-air recordings before deployment.
RF and text are paired at record level
Each released domain contains 200 million paired records. The counted unit is one bounded RF record with one or more text annotations. Individual complex-valued signal samples are not counted as dataset records.
The collection overview presents the current declared-scene qualification fixture and its evidence boundary. Review the RF-to-text qualification example before interpreting the pairing checks below.
| Domain | Coverage | Checks |
|---|---|---|
| Communications v1 | 200M paired RF-text records | Required RF and text fields present; annotation linked to the same RF record |
| Counter-UAS v1 | 200M paired RF-text records | Required RF and text fields present; annotation linked to the same RF record; surrogate signals identified in text |
These automated checks confirm that required fields exist and that each annotation points to the intended RF record. They do not judge whether every sentence is scientifically correct or prove that the annotation describes real-world RF accurately.
Evidence by dataset
Follow each result to its dataset definition, where the method, interpretation, limitation, and technical report are presented together.
01Stored-waveform physics: delay spread, mobility, and Doppler responseView figure02Communications scene realization and falsified occupancy assumptionView figure03Counter-UAS specification, decode, and separability checkpointsView figure04Counter-UAS scene and label realizationView figure05Counter-UAS synthetic-to-real transfer failureView figure