⚡ electrical engineering

Integrated Sensing, User Location and Orientation Estimation in RIS-Assisted Dynamic Rich Scattering Environment

This paper proposes a biLSTM-based adaptive framework that sequentially optimizes RIS configurations and beamforming vectors to accurately estimate user location and orientation in dynamic rich-scattering environments by progressively focusing pilot signals.

Original authors: Anum Umer, Ivo Müürsepp, Muhammad Mahtab Alam

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anum Umer, Ivo Müürsepp, Muhammad Mahtab Alam

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 in a completely dark, crowded room filled with people walking around, furniture, and mirrors. You are holding a flashlight (the User Equipment or UE), and you need to find exactly where you are standing and which way you are facing.

Normally, you would just look around. But in this room, the light bounces off the walls, the people, and the furniture so many times that it's impossible to tell where the original beam came from. It's like trying to find a specific echo in a canyon where a thousand other echoes are happening at once. This is what engineers call a "Rich Scattering Environment."

This paper proposes a clever solution using a high-tech "smart mirror wall" (called a RIS or Reconfigurable Intelligent Surface) and a very smart computer brain (an AI) to solve this puzzle.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Echo Chamber"

In a normal room, if you shout, you hear your voice come back clearly. In this "Rich Scattering" room, your shout hits a person, bounces to a chair, hits a mirror, bounces off a wall, and comes back to you mixed with thousands of other shouts.

The Challenge: The Base Station (the "listener" in the room) receives a messy jumble of signals. It doesn't know which part of the signal is the direct path and which part is a reflection. Traditional methods fail here because they assume the room is empty or simple.

2. The Hero: The "Smart Mirror Wall" (RIS)

Imagine the walls of the room are covered in thousands of tiny, magical mirrors. These aren't just static mirrors; they are programmable.

What they do: The Base Station can tell these tiny mirrors to change their angle instantly. One moment, a mirror might reflect light to the left; the next, it might reflect it to the right.
The Goal: By changing the mirrors, the Base Station can "tune" the room. It can try to focus the messy echoes into a single, clear beam pointing directly at you.

3. The Brain: The "Time-Traveling Detective" (biLSTM)

The paper uses a specific type of Artificial Intelligence called a biLSTM (Bidirectional Long Short-Term Memory).

The Analogy: Think of a detective trying to solve a crime. A normal detective looks at the evidence now. This AI detective is special because it looks at the evidence now, remembers what happened yesterday, and even predicts what might happen tomorrow.
How it helps: The AI watches the "messy echoes" over time. It learns that "Oh, when the mirrors are set to this pattern, the signal gets clearer." It learns the relationship between the moving people in the room and the signal quality. It essentially learns the "personality" of the room.

4. The Strategy: "Adaptive Sensing" (The Game of Clue)

Instead of guessing once and hoping for the best, the system plays a game of Clue (or Mastermind) over several turns:

Turn 1: You send a signal. The AI listens to the messy echo.
The AI's Move: Based on what it heard, the AI says, "Okay, the room is confusing. Let's change the mirrors to focus on the left side and tell the Base Station to listen more carefully there."
Turn 2: You send another signal. The AI listens again. It realizes, "Ah, there's a person walking near the left wall causing interference. Let's adjust the mirrors to avoid them."
The Result: With every turn, the AI gets smarter. It refines the mirror settings and the listening direction until the signal is perfectly focused on you.

5. The Two-Step Dance

The paper describes a two-part process happening simultaneously:

Step A: Mapping the Chaos: The AI uses a special part of its brain to figure out where the "Scattering Objects" (the moving people and furniture) are. It treats the moving people as obstacles it needs to map.
Step B: Pinpointing You: Using that map of the obstacles, the AI calculates exactly where you are and which way you are facing.

Why is this a big deal?

Old Way: Trying to find you in a crowded, dark room by guessing.
This Paper's Way: Using a smart, programmable room and a detective AI to actively "clean up" the noise, learn the layout of the room in real-time, and lock onto you with laser precision.

The Bottom Line

The researchers tested this in simulations with different numbers of moving people, different wall setups, and different amounts of "smart mirrors." They found that their AI-driven approach was much better than traditional mathematical formulas (which get confused by the chaos) and better than other AI methods that didn't try to map the moving obstacles.

In short: They built a system that doesn't just "listen" to a messy room; it actively "tunes" the room to make the signal clear, allowing it to find your location and orientation with incredible accuracy, even in the most chaotic environments.

1. Problem Statement

The paper addresses the challenge of User Equipment (UE) localization and orientation estimation in indoor Rich Scattering Environments (RSEs).

Context: Modern applications (6G, robotics, smart manufacturing) require precise positioning, but indoor environments are dominated by dense multipath propagation and dynamic changes caused by moving objects.
Challenge: In narrowband indoor settings, the UE pilot signal reflects off walls, Reconfigurable Intelligent Surfaces (RIS), and dynamic Scattering Objects (SOs). These overlapping paths with varying amplitudes and phases make it difficult to distinguish the direct path or extract standard parameters like Time-of-Arrival (ToA) or Angle-of-Arrival (AoA).
Goal: To jointly estimate the UE's 2D position ( $u$ ) and orientation ( $\alpha$ ) while simultaneously sensing the environment (estimating the locations of $M$ moving SOs) using a Multiple-Input-Multiple-Output (MIMO) system assisted by distributed RIS. The system must adaptively design RIS configurations and beamforming vectors over time to maximize localization accuracy.

2. Methodology

The authors propose a learning-based adaptive sensing framework that utilizes a Bidirectional Long Short-Term Memory (biLSTM) network to capture temporal dependencies in the channel measurements.

A. System Model

Setup: An indoor enclosure with a Base Station (BS), a multi-antenna UE, and multiple distributed RIS panels.
Dynamics: The environment contains $M$ randomly moving Scattering Objects (SOs).
Channel: Modeled using a physics-based end-to-end dipole array model that captures complex electromagnetic interactions (mutual coupling, polarization, scattering) rather than simple geometric ray tracing.
Operation: The UE transmits pilot signals. The BS receives these signals and uses a subset of RIS elements ( $S_{RIS}$ ) dedicated to sensing to estimate SO locations. Based on this, the system adaptively updates the RIS phase shifts, BS beamforming, and UE beamforming vectors for the next time slot.

B. Proposed Architecture (Two-Stage biLSTM)

The solution employs a neural network architecture with two distinct but interconnected biLSTM components:

Scattering Estimation biLSTM:
- Input: Magnitude and phase of signals received by the sensing RIS elements ( $S_{RIS}$ ).
- Function: Estimates the positions of the moving SOs ( $\hat{p}$ ) by learning the temporal evolution of the scattering environment.
UE Localization biLSTM:
- Input: Received pilot signals at the BS, the estimated SO positions from the first network, and the history of beamforming configurations.
- Function: Sequentially designs the RIS configuration vector ( $k$ ), BS beamforming vector ( $\omega_{BS}$ ), and UE beamforming vector ( $\omega_{UE}$ ) for the next time step.
- Output: Finally, at the end of the transmission sequence ( $T$ ), it outputs the estimated UE location ( $\hat{u}$ ) and orientation ( $\hat{\alpha}$ ).

C. Training and Optimization

Loss Function: The network is trained to minimize a composite loss function comprising the Root Mean Square Error (RMSE) of UE location, UE orientation, and SO location estimation.
Constraint Handling: The network outputs real and imaginary components which are normalized to satisfy the unit-modulus constraint (constant amplitude, variable phase) required for RIS elements and beamforming vectors.
Adaptive Strategy: Unlike greedy approaches that optimize for the immediate next step, this approach learns a sequence of actions over $T$ time frames to minimize the final localization error.

3. Analytical Benchmark

To validate the learning-based approach, the authors derived an analytical benchmark using the Bayesian Cramér-Rao Lower Bound (BCRLB).

This method treats the problem as an optimization task where the RIS configuration is chosen to minimize the conditional BCRLB of the UE position at each time step.
It serves as a theoretical lower bound and a comparison point to demonstrate the superiority of the data-driven approach in complex, non-linear RSEs.

4. Key Contributions

Joint Adaptive Design: A novel framework that jointly designs RIS configurations, BS beamforming, and UE beamforming vectors based on past measurements and environmental sensing, specifically for dynamic RSEs.
Scattering-Aware Localization: Unlike traditional methods that ignore scattering or treat it as noise, this approach explicitly models and estimates the dynamic scattering objects to improve localization accuracy.
BiLSTM Architecture: The use of biLSTM to capture long-range temporal dependencies in both the scattering environment and the received pilot signals, enabling the system to "learn" the evolving channel dynamics.
Comprehensive Evaluation: The study evaluates performance under diverse conditions, including:
- Different distributed RIS deployment strategies (1 to 4 walls).
- Varying numbers of moving SOs.
- Different numbers of sensing RIS elements.
- Comparison against BCRLB-based optimization and other learning baselines.

5. Simulation Results

The proposed approach was evaluated against several baselines: a beamforming-free biLSTM, a UE-localization-only biLSTM (ignoring scattering), and the BCRLB-based optimization.

Accuracy: The proposed method consistently achieved the lowest RMSE for both UE location and orientation estimation across all time frames ( $T$ ) and Signal-to-Noise Ratios (SNR).
Comparison with BCRLB: The BCRLB-based optimization performed significantly worse than the learning-based approach. The authors attribute this to the difficulty of analytically modeling the complex, non-linear interactions in a dynamic RSE, whereas the neural network learns these patterns directly from data.
Impact of RIS Deployment:
- Deploying RIS on all four walls yielded the best performance due to maximal spatial diversity.
- Single-wall deployments resulted in the highest error due to limited angular diversity.
- Increasing RIS density improved accuracy but showed diminishing returns once a certain density was reached.
Impact of Scattering Objects: As the number of moving SOs ( $M$ ) increased, the estimation error increased for all methods due to higher complexity and path coupling. However, the proposed method remained the most robust.
Sensing Elements: Increasing the number of dedicated sensing RIS elements ( $S_{RIS}$ ) significantly improved the accuracy of SO estimation, which in turn enhanced the overall localization performance.

6. Significance

This work demonstrates that adaptive sensing driven by deep learning is a viable and superior solution for 6G localization in complex indoor environments.

Robustness: It proves that learning-based methods can outperform theoretical optimization bounds (BCRLB) when the environment is too complex for analytical modeling.
Dual Functionality: It successfully integrates Integrated Sensing and Communication (ISAC), using the same infrastructure to both communicate and sense the environment (tracking moving objects) without requiring extra RF chains.
Practicality: The approach handles the non-linear, time-varying nature of real-world indoor scattering, offering a path toward reliable positioning for autonomous systems and smart manufacturing where GPS is unavailable.