⚡ electrical engineering

WiSER: A Wireless Scene Encoder for Geometry-Grounded Multi-View Wireless Prediction

This paper introduces WiSER, a geometry-grounded wireless scene encoder that leverages a shared sparse 3D scene representation to jointly predict dense radiomaps and variable-cardinality multipath channel impulse responses, outperforming single-task baselines and advancing the development of foundation models for AI-native wireless systems.

Original authors: Jing Qiao, Yiyang Guo, Hao Ye

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jing Qiao, Yiyang Guo, Hao Ye

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a radio signal (like your Wi-Fi or 5G) will behave inside a building. Traditionally, engineers have used two different "languages" to describe this:

The "Heat Map" Language: This shows a smooth, colorful picture of signal strength everywhere in the room (like a weather map showing rain).
The "Path List" Language: This is a specific list of individual signal bounces. It says, "The signal hit the wall, bounced to the table, and then hit the receiver," listing the time and strength of each specific bounce.

Until now, computer programs usually had to learn these two languages separately. If you wanted a heat map, you trained one AI. If you wanted a list of bounces, you trained a different AI. They didn't talk to each other, even though they were describing the exact same room.

WiSER is a new system that learns one single "mental map" of the room that can speak both languages at the same time.

Here is how it works, broken down into simple concepts:

1. The "Sparse Lego" Room

Instead of trying to process every single inch of empty air in a room (which is a waste of computer power), WiSER looks at the room like a set of sparse Lego blocks. It only keeps the blocks where there are actual objects—walls, furniture, doors—and ignores the empty space. This makes the computer's job much faster and more efficient.

2. The "Transmitter-Conditioned" Memory

Think of WiSER's brain as a smart librarian.

When you tell the librarian, "The radio tower is here," the librarian doesn't just look at the room; they look at the room specifically from the perspective of that tower.
This creates a "transmitter-conditioned memory." It's like the librarian rearranging the books on the shelves based on where you are standing, so the most relevant information is right in front of you.

3. The Two "Specialized Readers"

Once the librarian has this smart memory, two different "readers" can ask questions and get answers instantly, without needing to re-read the whole room:

Reader A (The Heat Map Artist): This reader looks at the librarian's memory and draws a smooth, detailed picture of signal strength for every spot on the floor. It uses a trick called "Ray-Corridor Gathering." Imagine shining a flashlight from the transmitter to a specific spot on the floor; this reader only pays attention to the objects in that narrow beam of light, ignoring the rest of the room to make a quick, accurate prediction.
Reader B (The Path Detective): This reader looks at the same memory but answers a different question: "What are the specific bounces?" It doesn't guess a fixed list; instead, it acts like a detective with a variable number of clues. It says, "I found 3 bounces for this spot," or "I found 5 bounces for that spot." It uses a special technique (called a DETR-style decoder) to figure out exactly how many bounces exist and what their timing and strength are, without forcing a fixed number of answers.

4. The "Training Gym"

To teach WiSER, the researchers didn't use real-world radio waves (which are messy and hard to control). Instead, they built a virtual gym using:

ScanNet++: A database of 3D scans of real rooms.
Sionna: A physics simulator that acts like a perfect, super-fast ray-tracing engine.

They used this gym to generate thousands of perfect "Heat Maps" and "Path Lists" for the same rooms. This allowed WiSER to learn that the "Heat Map" and the "Path List" are just two different ways of looking at the same physical reality.

The Results

When they tested WiSER, it proved that one brain can do two jobs better than two separate brains.

For Heat Maps: It predicted signal strength more accurately than previous methods that were designed only for heat maps.
For Path Lists: It predicted the timing and strength of signal bounces much better than previous methods that were designed only for path lists.

The Bottom Line

The paper claims that WiSER successfully built a reusable "scene encoder." Instead of training a new AI for every new room or every new type of wireless question, you can train this one system once. It learns the geometry of the room, and then you can ask it for a heat map or a list of bounces, and it will give you a high-quality answer for both.

The researchers also found that the best way to teach this system was not to throw all the data at it at once, but to teach it the heat map job first, then the path list job, and then let them practice together in turns (a "warm-started alternating" strategy). This prevented the system from getting confused by the two different types of tasks.

Technical Summary: WiSER - A Wireless Scene Encoder for Geometry-Grounded Multi-View Wireless Prediction

Problem Formulation
Indoor wireless propagation is a scene-conditioned physical process determined by 3D geometry, radio-material properties, and transmitter-receiver configurations. While classical stochastic models abstract away site-specific structures and deterministic ray-tracing simulators are computationally expensive for repeated queries, learning-based methods have emerged as surrogates. However, existing learning-based approaches are typically siloed: radiomap predictors focus on dense spatial field regression (aggregate coverage), while Channel Impulse Response (CIR) predictors focus on link-level multipath structures (sparse, unordered sets of delay-power taps). This separation raises the question of whether heterogeneous wireless views can be decoded from a shared, reusable representation of the same physical scene. Current datasets often lack the co-registration required to train such a unified model, as they typically target single output representations (e.g., fixed channel matrices or standalone radiomaps) rather than aligned multi-view supervision.

Methodology: The WiSER Architecture
The paper proposes WiSER (Wireless Scene Encoder), a framework designed to learn a reusable, transmitter-conditioned sparse 3D scene representation that supports both dense radiomap prediction and variable-cardinality multipath CIR prediction.

Scene Representation: The indoor environment is represented as a sparse 3D voxel set ( $V_S$ ), retaining only occupied surfaces and objects to avoid dense computation over empty space. Features include occupancy, color, semantic category, and material attributes.
Transmitter-Conditioned Sparse Scene Memory: A shared encoder maps the sparse voxel scene and a transmitter location ( $x_t$ $x_{t}$ ) into a multi-resolution memory ( $Z_\phi$ $Z_{ϕ}$ ).
- Local Memories ( $X_{loc}$ ): Fine-scale memories retain geometry near occupied surfaces and remain transmitter-agnostic to ensure reusability across different transmitter placements.
- Global Memory ( $M_{tx}$ ): A compact global memory summarizes the scene under the current transmitter placement. It is modulated by an embedding of $x_t$ (derived from Fourier positional features and an MLP) using Adaptive Layer Normalization zero (AdaLN-zero) conditioning.
Dual Decoders: Two structure-aware decoders query the shared memory:
- Ray-Corridor Radiomap Decoder: For dense receiver-plane path-gain prediction, this decoder employs a "ray-corridor gathering" mechanism. Instead of attending to all voxels, it selects a compact set of scene tokens near the transmitter-receiver segment and endpoints. This acts as a geometry-constrained attention sparsifier, injecting a weak propagation prior by exposing queries to likely blockers and reflectors. The output is a dense path-gain field.
- DETR-Style CIR Set Decoder: For multipath CIR prediction, this decoder treats the output as an unordered set of delay-power taps with variable cardinality. It uses a Detection Transformer (DETR) style head with $K$ learnable candidate tap queries. The queries are conditioned on the receiver location and attend to the global scene memory. A Hungarian matching algorithm aligns predicted taps with ground-truth paths, allowing for permutation-invariant supervision.
Training Strategy: Due to the heterogeneity of the tasks (dense regression vs. sparse set prediction) and differing loss scales, the paper employs a warm-started alternating optimization strategy. The model is initialized with single-task checkpoints, followed by a short joint alignment phase and alternating task-focused updates where the shared encoder and the active decoder are updated while the inactive decoder is frozen.

Dataset Construction
To enable this multi-view learning, the authors constructed a co-registered indoor scene and wireless dataset pipeline using ScanNet++ scenes and Sionna Ray Tracing.

Co-Registration: The pipeline converts a single ScanNet++ scene into both a sparse voxel representation for the model and a Sionna-compatible radio scene for label generation.
Labels: It generates aligned dense radiomap labels (path-gain fields) and unordered multipath CIR tap sets (delay and power) under a common coordinate frame and propagation configuration.
Multipath Consolidation: To match the resolution of the sparse voxel input, raw ray-tracing paths are consolidated into representative taps based on interaction depth, interaction-type sequences, and quantized bounce-voxel sequences, ensuring the supervision targets are resolvable by the scene representation.

Key Contributions

Problem Formulation: The paper formulates radiomap and multipath CIR prediction as complementary views of the same scene-conditioned propagation process, establishing a concrete setting for studying reusable geometry-grounded representations.
Dataset Pipeline: It introduces a ScanNet++/Sionna-based pipeline that co-registers sparse 3D scene geometry, transmitter/receiver configurations, and both dense radiomap and path-level CIR labels.
WiSER Model: It proposes the WiSER architecture, featuring a transmitter-conditioned sparse 3D encoder and two specialized decoders (ray-corridor for radiomaps and DETR-style for CIRs).
Empirical Validation: It provides experimental evidence that a single shared representation can support both aggregate coverage and path-level channel prediction, outperforming scene-specific baselines.

Experimental Results
The model was evaluated on a 100-scene training set and 10 held-out scenes.

Radiomap Prediction: WiSER achieved a Mean Absolute Error (MAE) of 3.834 dB and RMSE of 5.500 dB, outperforming scene-specific baselines like NeRF2 (8.238 dB MAE) and RF-3DGS (4.585 dB MAE). Notably, WiSER was trained once on the multi-scene set and reused, whereas baselines were optimized separately for each scene.
Multipath CIR Prediction: WiSER achieved a matched peak-power MAE of 5.89 dB and delay MAE of 0.61 ns, significantly outperforming reference baselines including a 3D CNN (11.50 dB peak-power MAE, 1.50 ns delay MAE).
Ablation Studies: Removing the ray-corridor gathering or receiver self-attention degraded radiomap performance. Similarly, removing cross-attention or self-attention in the CIR decoder severely degraded CIR accuracy, particularly in path-count prediction.
Training Strategy: Direct joint training from random initialization resulted in underfitting. The warm-started alternating strategy was shown to be necessary for effectively training the shared encoder under heterogeneous supervision.

Significance and Claims
The paper claims that WiSER demonstrates that transmitter-conditioned sparse 3D scene representations can serve as reusable wireless scene encoders for heterogeneous propagation queries. By learning a shared scene memory rather than isolated predictors, the model supports both dense coverage-level and sparse path-level predictions within a single architecture. The authors position this as a "geometry-grounded step toward representation learning and foundation-model development for AI-native wireless systems," suggesting that the physical environment itself can provide a reusable representation for multiple downstream wireless tasks, provided the decoders are matched to the specific structure of each output space.