Hybrid Neural Network-Based Indoor Localisation System for Mobile Robots Using CSI Data in a Robotics Simulator

AI summary

Overview

Research area: Indoor localisation for mobile robots using wireless Channel State Information (CSI) and deep learning, evaluated within a robotics simulator.

Technical level: Intermediate. Readers should be comfortable with neural network architectures (CNN, MLP), wireless signal concepts like CSI and MIMO, and basic mobile robotics (ROS, Kalman filters). The paper is applied engineering rather than deep theory.

Scope: The paper develops and simulates a hybrid CNN+MLP model that estimates a mobile robot's 2D position from CSI readings, then evaluates it inside Webots with ROS across three navigation scenarios, noise levels, and antenna counts.

What This Paper Is About

Mobile robots need accurate position knowledge to navigate, but GPS and visual SLAM often fail indoors. Wireless CSI from Massive MIMO systems offers an alternative signal source, yet collecting real fingerprinting data is expensive and hard to validate at scale. The paper addresses this by training a hybrid neural network on an existing CSI dataset, converting the signals into synthetic images, and testing the resulting localisation model inside a robotics simulator to see how well it holds up under motion, noise, and disruption.

Key Contributions

1. A Hybrid Neural Network (HyNN) for CSI-based robot localisation. The model joins a Convolutional Neural Network branch (processing synthetic CSI images) with a Multilayer Perceptron branch (processing raw CSI-derived features) to output 2D robot positions.

2. A pipeline that converts CSI readings into image inputs. Using the TINTO tool, complex CSI matrices are transformed into 35×35 pixel images with blurring applied, enabling CNNs to learn spatial patterns from wireless signal data.

3. Integration of the localisation model with a robotics simulator and ROS. The scenario is reconstructed in Webots, the e-puck2 robot follows predefined paths, and the HyNN's predictions are fed back into the simulated robot loop for realistic evaluation.

4. Coupling the HyNN with a Kalman filter state estimator. The paper analyses when a simple 2D Kalman filter improves localisation accuracy and when it degrades it, particularly under high noise or abrupt displacement.

Main Findings

• More antennas mean better accuracy. Test-set mean error drops from 164.98 mm with 8 antennas to 49.66 mm with 64 antennas in the ULA configuration, reflecting the spatial diversity of larger MIMO arrays.

• Prediction speed falls as antenna count rises. Models achieve roughly 5, 4, 4, and 3 predictions per second for 8, 16, 32, and 64 antennas respectively, which sets the practical frame rate for evaluation.

• Kalman filtering sharply reduces error under noise. In Experiment 1 (uniform motion), the 8-antenna configuration with high noise went from 381.23 mm without the filter to 272.25 mm with it; the 64-antenna case improved from 45.20 mm to 25.80 mm.

• The Kalman filter can hurt when predictions are already good. With 64 antennas and low noise, or during the kidnapped robot scenario, the simplified constant-velocity filter lagged behind the HyNN and increased mean error — the 64-antenna, no-noise kidnapping case rose from 35.56 mm to 49.44 mm.

• Obstacle avoidance exposes the filter's limits. In Experiment 2, variable robot speed broke the constant-velocity assumption, so the Kalman filter helped mostly in high-noise, low-antenna settings and offered little benefit at 64 antennas.

• The kidnapped robot problem is a weak point. After a forced displacement, the filter required a long convergence period because robot motion became highly nonlinear; the HyNN alone handled the recovery better when its baseline accuracy was high.

• 32 or 64 antennas meet a practical accuracy bar. The authors conclude that these configurations deliver average errors around 10 cm or less, which they compare favourably to the robot's physical size.

Methodology in Plain English

The researchers started from an existing CSI dataset recorded with 64 antennas and 100 subcarriers over 252,004 positions spaced 5 mm apart in a small indoor area. They picked the Uniform Linear Array (ULA) configuration because it is easy to replicate physically, and also built reduced datasets with 8, 16, and 32 antennas to test how array size affects accuracy.

Raw CSI values are complex numbers, so they were split into modulus and argument (polar form) and stored alongside the corresponding X and Y coordinates. For the CNN branch, the TINTO tool converted each CSI reading into a small 35×35 pixel image using dimensionality reduction, coordinate scaling, and blurring. Two separate models were trained per configuration — one for X, one for Y — using an 85/10/5 train-validation-test split in Keras and TensorFlow.

For evaluation, the real indoor space was rebuilt in the Webots simulator with coordinates matching the dataset. An e-puck2 robot traversed predefined routes while a Supervisor node recorded its true position. The CSI readings used during testing were computed as the mean of at least nine nearby dataset positions to avoid reusing training samples, then subsampled to match each model's real prediction rate and corrupted with Gaussian noise at 0%, 10%, 20%, and 30% levels. A simple 2D Kalman filter tracking position and velocity was optionally applied on top of the HyNN outputs. Three experiments were run: straight-line uniform motion, obstacle avoidance, and a kidnapped robot scenario where the robot is abruptly teleported mid-route.

Why This Matters

Research impact. The work offers a reusable procedure for validating wireless localisation algorithms without expensive real-world fingerprinting campaigns. By combining an existing dataset, a signal-to-image transformation tool, a neural network, and a robotics simulator, it shows how simulation can stand in for costly physical trials and how state estimators behave when layered on top of learned localisation.

Real-world applications:

• Warehouse and logistics robots that must localise precisely where GPS is unavailable and ceiling-mounted infrastructure is already present.
• Hospital service robots navigating corridors and wards where visual features are repetitive or dynamically occluded.
• Industrial inspection robots operating in cluttered indoor plants with metal surfaces that degrade other sensing modalities.
• Assistive or telepresence robots in offices and homes, where Wi-Fi access points already exist and CSI could be harvested opportunistically.

Industry relevance. The finding that 32–64 antennas yield roughly 10 cm accuracy gives infrastructure planners a concrete benchmark. The observation that simple Kalman filters can degrade performance when the underlying model is already accurate is directly actionable for teams deploying hybrid learning-plus-filtering stacks. The public GitHub repository also lowers the barrier for companies wanting to prototype similar systems.

Future Directions

• Real-world validation. The authors plan to test on physical robots in diverse indoor spaces, which is essential because the entire evaluation here was simulation-based and depended on synthetic image generation and injected Gaussian noise.

• Better state estimators. Monte Carlo localisation and other filters suited to nonlinear, dynamic contexts could address the Kalman filter's poor showing during the kidnapped robot scenario and obstacle avoidance.

• Cost versus accuracy trade-offs. Studying configurations with fewer antennas and different wireless datasets would clarify where the accuracy ceiling sits and how much infrastructure is genuinely needed.

• Generalisation beyond the studied dataset. The paper claims its procedure is transferable, but it was only demonstrated on one ULA dataset; testing on URA, DIS, and datasets from other buildings would substantiate that claim.

Target Audience

Researchers and graduate students working on indoor positioning, wireless sensing, or applied deep learning for robotics. Robotics engineers evaluating whether CSI-based localisation is viable for their deployments will find the antenna-count and noise results directly useful. Simulation specialists interested in coupling machine learning models to ROS and Webots will also benefit, as will practitioners who want a template for validating localisation algorithms before committing to field trials.

Authors’ abstract

We present a hybrid neural network model for inferring the position of mobile robots using Channel State Information (CSI) data from a Massive MIMO system. By leveraging an existing CSI dataset, our approach integrates a Convolutional Neural Network (CNN) with a Multilayer Perceptron (MLP) to form a Hybrid Neural Network (HyNN) that estimates 2D robot positions. CSI readings are converted into synthetic images using the TINTO tool. The localisation solution is integrated with a robotics simulator, and the Robot Operating System (ROS), which facilitates its evaluation through heterogeneous test cases, and the adoption of state estimators like Kalman filters. Our contributions illustrate the potential of our HyNN model in achieving precise indoor localisation and navigation for mobile robots in complex environments. The study follows, and proposes, a generalisable procedure applicable beyond the specific use case studied, making it adaptable to different scenarios and datasets.