Abstract
Indoor localisation remains a significant challenge for autonomous robotics, particularly in environments where GPS is unavailable. Traditional methods like SLAM often suffer from cumulative drift, while hardware-based solutions require expensive environmental modifications. This paper presents an efficient neural mapping approach that treats localisation as a 2D regression task. By training compact Convolutional Neural Networks (CNNs) on massive datasets of up to 200,000 samples, we achieve centimetre-level accuracy. The system enables real-time absolute pose estimation from a single RGB camera on embedded hardware, demonstrated through successful navigation in a real-world apartment.
Introduction
Localising a robot indoors without external infrastructure is a "hard" problem in robotics. While Visual SLAM is a popular solution, its reliance on relative motion estimation leads to "drift" over time. Neural mapping offers an alternative by learning an implicit map of a scene, allowing a network to predict an absolute pose $(x, y, \theta)$ from a single image. However, previous attempts have often been too computationally intensive for mobile deployment or lacked the precision required for tight indoor navigation.
Methodology
The proposed system treats localisation as a regression task rather than a classification problem. A CNN backbone is modified to output three continuous values representing the 2D coordinates and the yaw (orientation) of the robot.
The 2D Constraint
A key innovation is constraining the global localisation problem to 3 Degrees of Freedom (3-DOF: $x, y, \theta$). Unlike prior work focusing on 6-DOF (3D position and orientation), this constraint significantly reduces the search space and model complexity, making it suitable for ground-based vehicles.
Architecture and Training
Several architectures were benchmarked, including ResNet, VGG19, and MobileNet-V2. Training utilised L1 loss, which proved more effective than L2 for penalising small-magnitude errors without over-penalising outliers. To handle the massive datasets required for convergence, a "learning rate reset" strategy was employed to help the model escape local minima during long training sessions.
Datasets
The scale of data is a core contribution of this work. While typical neural mapping studies use a few thousand images, this research leverages hundreds of thousands of samples across simulated and real environments.
- Simulated Environments: Three photorealistic Unity scenes—a Cabin ($105m^2$), a Warehouse ($1200m^2$), and a Temple ($4200m^2$)—each provided 200,000 training samples.
- Real-World Environment: A $64m^2$ studio apartment where 32,638 samples were collected using a Clearpath Husky UGV. Ground truth was established using an ultrasonic positioning system accurate to $\pm2cm$.
Results and Insights
Accuracy and Performance
In simulation, the models achieved a mean position error of less than 4cm. In the real-world apartment, the mean accuracy was approximately 9cm. The system demonstrated high inference speeds across different hardware:
- Desktop (RTX 3090): 220 FPS
- Embedded GPU (Jetson Xavier): 35 FPS
- Onboard CPU (Intel i5): 6 FPS
Key Findings
- Data Volume: Accuracy gains continued up to 200,000 samples, especially in larger environments like the Warehouse.
- Resolution Paradox: Lower resolution (128x128) outperformed higher resolutions (512x512). Lower resolutions likely force the network to focus on global geometric context rather than local textures or noise.
- Absolute vs. Relative: Because the method provides an absolute location, errors do not compound over time, offering a significant advantage over SLAM for long-term autonomy.
Conclusion
By combining massive datasets with a constrained 2D pose regression task, we demonstrate that neural mapping can provide precise, real-time indoor localisation for UGVs. Future work will address challenges such as motion blur, dynamic scenes, and the need for environment-specific retraining through transfer learning.
Summary updated