Introduction
In complex indoor environments such as warehouses or search-and-rescue zones, robots often need to locate a specific target, such as a person with a mobile device or a wireless beacon. While traditional navigation systems rely heavily on visual sensors like cameras and LiDAR, these are limited by line-of-sight (LOS) requirements. This paper introduces a multi-modal navigation framework that enables robots to "listen" to millimeter-wave (mmWave) signals to find targets even when they are obscured by walls or furniture.
The core challenge in wireless navigation is the "multipath" problem: signals bounce off surfaces, meaning the strongest received signal might not point directly toward the target. By employing a machine-learning classifier to distinguish between direct paths and complex reflections, the proposed system allows a robot to intelligently decide when to trust a signal's direction, significantly outperforming vision-only baselines.
Methodology
The proposed framework integrates a three-stage wireless processing pipeline into an Active Neural-SLAM module.
1. Path Estimation
The robot first processes received mmWave signals using a low-rank tensor decomposition method. This technique identifies multiple signal paths and estimates critical parameters, including Angle of Arrival (AoA), Angle of Departure (AoD), relative delay, and Signal-to-Noise Ratio (SNR).
2. Link State Classification
A two-layer fully connected neural network is used to classify the "link state" of the received signals. Distinguishing between these states is vital for determining the reliability of the directional data.
The classifier categorizes signals into four states:
- LOS (Line-of-Sight): The direct path to the target.
- First-order NLOS: A signal that has bounced once off a surface.
- Higher-order NLOS: A signal that has undergone multiple reflections.
- Outage: No usable signal is detected.
3. Intelligent Navigation Logic
The robot follows a strategic decision-making process: if the classifier identifies a LOS or First-order NLOS path, the robot trusts the estimated angle and moves toward it. If higher-order reflections or an outage are detected, the robot ignores the wireless data and reverts to standard visual exploration to map the environment and avoid obstacles.
Dataset and Simulation
To train and validate the system, the researchers developed a novel multi-modal dataset by augmenting the Gibson 3D indoor dataset. They used the Remcom Wireless InSite package to perform high-fidelity ray tracing, simulating 28 GHz mmWave propagation across 38 unique indoor environments. This resulted in a unified dataset combining 3D camera data, LiDAR scans, and wireless path estimates, which has been released publicly to support further research.
Results and Discussion
The experimental results demonstrate the robustness of the link state classification approach. The neural network achieved an 88% accuracy in identifying link states in entirely new, unseen environments, suggesting that the physics of signal reflection are consistent enough for generalizable learning.
In terms of navigation performance, the wireless-assisted method achieved a success rate of approximately 80% in "Hard" difficulty scenarios. In contrast, a baseline that simply followed the strongest signal (SNR-based) dropped to nearly 60% success because it frequently pursued "ghost" reflections. Furthermore, the proposed method allowed the robot to reach targets with efficiency levels approaching an "oracle" baseline that possesses perfect target location knowledge.
Conclusion and Future Work
This research highlights that signal reflections are not merely noise but can be valuable navigational cues if classified correctly. By combining the global guidance of wireless signals with the local obstacle avoidance of visual SLAM, robots can navigate complex spaces more effectively.
Future work will address current limitations, such as the 2D simplification of the navigation logic. Extending the framework to 3D will be essential for aerial robotics. Additionally, moving beyond separate processing to joint sensor fusion (e.g., via transformer networks) could further enhance robustness. Finally, implementing these sophisticated phased-array antenna techniques on low-cost, power-constrained hardware remains a key engineering challenge for real-world deployment.
Summary updated