Fast and Scalable Signal Inference for Active Robotic Source Seeking

Christopher E. Denniston, Oriana Peltzer, Joshua Ott, Sangwoo Moon, Sung-Kyun Kim, Gaurav S. Sukhatme, Mykel J. Kochenderfer, Mac Schwager, Ali-akbar Agha-mohammadi

Introduction

In search-and-rescue or disaster response scenarios, locating individuals via wireless signals (e.g., radio, Bluetooth, Wi-Fi) is a critical task, especially when visual information is obscured by smoke or debris. This process, known as Active Source Seeking, involves a robot navigating an unknown environment to identify the strongest signal source.

Traditionally, Gaussian Processes (GPs) have been the standard for modeling these signals. However, GPs face significant challenges in real-world deployments:

  1. Scalability: Computational complexity grows cubically with the number of measurements, leading to slower processing as exploration progresses.
  2. Obstacle Blindness: Standard GP kernels often fail to account for physical barriers, such as concrete walls, which significantly attenuate signals.
Overview of the hierarchical signal inference system, showing global and local signal beliefs alongside traversability maps.

Methodology: Hierarchical Factor Graphs

To overcome these limitations, we propose a hierarchical factor graph approach that splits signal inference into two distinct components:

Global Factor Graph

The global graph maintains a sparse "skeleton" of signal strength across the entire explored area. It utilizes "link factors" to represent spatial correlation, providing a computationally efficient alternative to the dense covariance matrices used in GPs.

Local Factor Graph

A high-resolution grid centered on the robot incorporates real-time occupancy information. By increasing signal uncertainty in areas with obstacles, the model reflects physical phenomena like attenuation and reflection without requiring complex wave physics modeling. This encourages the robot to "peek" around corners to resolve uncertainty.

Integrated Planning

The system employs the PLGRIM (Probabilistic Local and Global Information-based Roadmap) planner to balance local exploration (investigating nearby high-uncertainty areas) and global planning (targeting distant frontiers likely to contain the source).

Experimental Setup

The system was validated through both simulation and real-world hardware trials:

  • Simulation: Data was collected from a robot mapping signal-to-noise ratios (SNR) in a parking garage environment.
  • Hardware: The approach was deployed on a Boston Dynamics Spot robot navigating a complex hallway environment to locate a hidden radio source.

Key Results

The Factor Graph approach demonstrated clear advantages over traditional GP methods in both efficiency and performance:

  • Superior Signal Acquisition: In simulations, the proposed method consistently found higher maximum signal strengths and achieved higher cumulative signal totals compared to standard and limited GPs.
  • Constant-Time Inference: Unlike GPs, which saw exponential increases in solve time as more data was added, the Factor Graph maintained low, stable inference times. A hash-table-based caching system allowed for near-instantaneous signal belief retrieval when the robot's state remained relatively static.
Performance comparison between the proposed Factor Graph approach and Gaussian Process (GP) methods across two radio sources.

Conclusion and Future Work

By framing signal estimation as a sparse optimization problem, this work enables fast, scalable, and obstacle-aware source seeking. Future research will focus on multi-source environments, time-varying signals, and extending the model to multi-robot swarms for collaborative search operations.

Summary updated

BibTeX
@inproceedings{denniston2023fast,
  author = {Christopher E. Denniston and Oriana Peltzer and Joshua Ott and Sangwoo Moon and Sung-Kyun Kim and Gaurav S. Sukhatme and Mykel J. Kochenderfer and Mac Schwager and Ali-akbar Agha-mohammadi},
  title = {{Fast and Scalable Signal Inference for Active Robotic Source Seeking}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10161445},
}
Figure