[Paper Review] A Robot Web for Distributed Many-Device Localisation

Riku Murai, Joseph Ortiz+3 more|arXiv (Cornell University)|Feb 7, 2022

Energy Efficient Wireless Sensor Networks51 references4 citations

TL;DR

This paper proposes Robot Web, a distributed, scalable framework for many-device localisation using Gaussian Belief Propagation (GBP) on a shared non-linear factor graph, enabling accurate global localisation via asynchronous, peer-to-peer communication without centralised coordination. The system achieves accuracy comparable to batch optimisation while maintaining high computational and communication efficiency, even under high outlier rates and communication failures.

ABSTRACT

We show that a distributed network of robots or other devices which make measurements of each other can collaborate to globally localise via efficient ad-hoc peer to peer communication. Our Robot Web solution is based on Gaussian Belief Propagation on the fundamental non-linear factor graph describing the probabilistic structure of all of the observations robots make internally or of each other, and is flexible for any type of robot, motion or sensor. We define a simple and efficient communication protocol which can be implemented by the publishing and reading of web pages or other asynchronous communication technologies. We show in simulations with up to 1000 robots interacting in arbitrary patterns that our solution convergently achieves global accuracy as accurate as a centralised non-linear factor graph solver while operating with high distributed efficiency of computation and communication. Via the use of robust factors in GBP, our method is tolerant to a high percentage of faults in sensor measurements or dropped communication packets.

Motivation & Objective

• To enable fully distributed, scalable localisation for large numbers of heterogeneous robots or devices without relying on a centralised system or cloud infrastructure.
• To address the challenge of global localisation in dynamic, asynchronous, and unreliable communication environments common in real-world multi-robot systems.
• To develop a system that maintains high accuracy despite faulty sensors, communication dropouts, and changing network topologies.
• To demonstrate feasibility on real robots with limited onboard computational resources, ensuring practical deployability.
• To establish a foundation for a future open, interoperable 'Robot Web' akin to the World Wide Web, enabling open exchange of geometric and spatial information across devices.

Proposed method

• The system models multi-robot localisation as a non-linear factor graph, where each robot maintains its own local fragment of the global graph, including pose variables, odometry factors, and inter-robot measurement factors.
• Gaussian Belief Propagation (GBP) is used as the inference engine, enabling distributed, iterative message passing across robots to converge to globally consistent localisation estimates.
• Each robot publishes and updates a 'Robot Web Page'—a representation of outgoing messages—using standardised, asynchronous communication protocols such as web pages or other lightweight data exchange mechanisms.
• The method is extended to support Lie Groups for accurate representation of 3D rotations and rigid-body transformations, critical for practical robotics applications.
• Robust factors are integrated into GBP to reject outlier measurements caused by faulty sensors or noisy observations, enhancing resilience to non-Gaussian noise.
• A time-windowing mechanism is applied to limit storage and communication overhead by discarding old pose variables and factors beyond a threshold, enabling long-term operation.

Experimental results

Research questions

• RQ1 Can a fully distributed, peer-to-peer system achieve global localisation accuracy comparable to centralised batch optimisation in large-scale multi-robot systems?
• RQ2 How well does the system perform under high levels of communication failure, sensor noise, and outlier measurements?
• RQ3 Can the system maintain scalability and convergence when robots dynamically join or leave the network, or reconfigure sensors in real time?
• RQ4 Is it feasible to implement such a system on real robots with limited onboard computational resources?
• RQ5 Can the system be generalised to support arbitrary robot types, motion models, and sensor modalities without requiring prior coordination or topology knowledge?

Key findings

• The Robot Web system achieves global localisation accuracy that matches that of a centralised non-linear factor graph solver in simulations with up to 1000 robots.
• The system demonstrates high robustness, successfully rejecting a high percentage of outlier measurements and maintaining convergence even under significant communication failures.
• In real-world experiments, the system successfully localised nine physical robots using only onboard computational resources, confirming practical feasibility.
• The use of time-windowing effectively bounds storage and communication overhead, enabling long-term operation without system degradation.
• The system is fully dynamic and supports arbitrary robot topologies, including disjoint connectivity and online changes in sensor configuration.
• The integration of robust factors into GBP significantly improves resilience to non-Gaussian noise, outperforming standard GBP in noisy environments.