CHANNELS4WORLD: An Environment-Scale Dataset for Wireless Foundation Models
Organizations: School of Integrated Technology, Yonsei University, Seoul 03722, South Korea · School of Electrical and Electronic Engineering, Yonsei University, Seoul 03722, South Korea · NVIDIA, 92400 Courbevoie, France · Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA 92093 USA
Abstract
Wireless foundation models learn reusable representations across tasks, configurations, and deployment environments. Their pretraining requires environmental breadth, yet wireless datasets often sample few sites densely. We introduce Channels4World, a scene-centered dataset of 100,000 distinct outdoor RF scenes and a procedural method for generating them. Each scene links a generated environment and RF material assignments to 32 base-station (BS) positions, 16,384 UE positions, and aligned variable-length multipath and ordered interaction records at 3.5 and 7~GHz. The corpus comprises 52.43 billion physical BS-UE pairs, 104.86 billion frequency-specific evaluations, and a 29.14-TB packaged payload. Scene identity aligns geometry, placements, channels, and interactions across frequencies, allowing environmental breadth and within-scene observation density to vary independently. The public release offers predefined subsets at multiple scales and separate access to corpus components. A deterministic 90K/10K scene-level train-test split keeps both frequencies and all scene components together. We characterize morphology, RF materials, and transceiver placement across three generation profiles. Controlled propagation and path-retention analyses show that diffraction materially affects retained-link availability and multipath structure, while the retention policy reduces storage substantially with little loss in represented aggregate power. Under matched training exposure, models trained on broader scene collections perform better on held-out scenes, although an equal-episode control shows that scene count alone cannot explain the gain. By making environmental breadth explicit and controllable, Channels4World supports the development and evaluation of wireless foundation models for generalization across deployment environments.
Figures & tables
| Resource | Environment source | Environment inventory / unit | Within-environment sampling | Native RF output | Aligned environment data | Resource availability |
|---|---|---|---|---|---|---|
| DeepMIMO v4 [ 3 , 10 ] | Mixed site-specific ray-tracing scenarios | 200+ scenarios in the v4 database; site extent, carrier, and deployment vary | Scenario-specific Tx/Rx sets; counts and spatial sampling vary by scenario | Per-path ray-tracing data; configurable MIMO channels computed on demand | Tx/Rx positions and path data; geometry, materials, and system metadata where provided | Public v4 toolchain and scenario database |
| WAIR-D [ 5 ] | Geographic import (OSM, 40+ cities) | 10,000 sparse environments; 100 dense environments | Sparse: 5 BS 30 UE. Dense: 1 BS 10,000 UE | Ray-traced channels | Environment image, BS/UE positions, and region size | Public dataset |
| RadioMapSeer [ 11 ] | Geographic import (OSM, 6 cities) | 701 maps, each m | 80 Tx/map on a 1-m receiver grid | Path-loss radio maps | OSM city maps, roads, Tx locations, and simulator variants | Public dataset |
| U6G XL-MIMO [ 12 ] | Geographic import (OSM, Nanjing) | 800 scenes, each km | 98 Tx configurations/scene on a receiver grid at 10-m spacing | 78,400 configuration-specific beamformed received-power maps | Height and beam maps, separately available meshes, and fixed material roles: marble buildings, metallic roofs, and concrete ground | Public data and code |
| Wi3R / WiPTR [ 13 ] | Procedural generation (indoor) | 5,000 / 12,000 scenes | 5 Tx 200 Rx / 1–15 Tx and up to 200 Rx | Per-path channels (25 strongest paths) | 3D meshes, fixed role-based materials (2 / 6 classes), and transceiver placements | Research-use release |
| OpenPathNet [ 14 ] | Geographic import (OSM) with procedural fallback | Paper: 12,000 m scenes. Releases: 1,700 curated / 10,000 larger | 1 Tx on a receiver grid at 1-m spacing | Per-path gain, delay, angles, and type | Scene XML and meshes with assigned materials, placements, and path records; scene-level material variation not characterized | Public generator and two releases |
| Unit | Key | Meaning |
|---|---|---|
| BS–UE pair | One BS–UE position pair in scene , shared by both frequencies | |
| Frequency-specific evaluation | That pair evaluated at one frequency | |
| Retained link | with | An evaluation with at least one retained path |
| Retained path | The th power-ordered path of a retained link |
| Property | Fixed setting |
|---|---|
| Direction and carriers | Downlink at 3.5 and 7 GHz |
| Antenna convention | Single-element isotropic V-polarized transmit and receive antennas |
| Modeled mechanisms | Line of sight (LoS), specular reflection, and diffraction with the lit region enabled |
| Mechanisms not modeled | Diffuse scattering, refraction or transmission, and edge diffraction |
| Maximum interaction depth | 4 |
| Material model | ITU-R P.2040-3 classes with frequency-dependent properties |
| Profile | Frequency | Retained links (bil.) | Retention (%) | Retained paths (bil.) | Paths / retained link |
| Urban | 3.5 GHz | 23.35 | 89.07 | 257.78 | 11.04 |
| Urban | 7 GHz | 21.13 | 80.61 | 188.93 | 8.94 |
| Metropolitan | 3.5 GHz | 9.02 | 68.84 | 87.81 | 9.73 |
| Metropolitan | 7 GHz | 7.42 | 56.58 | 62.75 | 8.46 |
| Suburban | 3.5 GHz | 12.40 | 94.63 | 119.70 | 9.65 |
| Suburban | 7 GHz | 12.00 | 91.52 | 83.91 | 7.00 |
| Policy | Logical storage | vs. corpus policy |
|---|---|---|
| dB (corpus policy) | 51.46 GB | 1.00 |
| dB | 78.15 GB | 1.52 |
| , threshold disabled | 81.42 GB | 1.58 |
| , threshold disabled | 206.40 GB | 4.01 |
| Block | Configuration |
|---|---|
| Shared attention / feed-forward | Width 384; 6 heads; feed-forward width 1536 |
| Context within-link attention | 4 self-attention layers |
| Environment aggregation | 32 latent tokens; 6 latent cross/self-attention pairs; 5 reciprocal context cross/self-attention pairs |
| Query–environment interaction | 8 bidirectional cross-attention stages |
| Query within-link post-attention | 2 self-attention layers |
| Prediction heads | Two 768-wide GELU hidden layers; separate path-mixture logit, mean, and scale heads |
| Property | Setting |
|---|---|
| Carrier / learning view | 7 GHz / learning view excludes diffraction-flagged retained paths |
| Main training tiers | 100 / 1K / 10K / 90K scenes |
| Equal-episode control | E1K: 1K scenes, 32 eligible BS episodes/scene; E32K-BS1: 32K scenes, one fixed BS episode/scene; 32K available episodes each |
| Episode | One eligible scene–BS pair; 192 context and 32 disjoint LoS-eligible query links; up to 16 retained paths/link |
| Model / output | 83.50M parameters; 32 environment tokens; eight-component position GMMs and relative-gain prediction |
| Effective global batch | 192 BS-centered episodes per optimizer step |
Wireless Communications
Wireless communications studies how information is sent over radio links and how networks manage signals, spectrum, and resources reliably.
See all 432 papers