💻 computer science

Probabilistic Multi-Robot Gas Source Localization with Uncalibrated Sensors: A Distributed Estimation Approach

This paper presents a distributed probabilistic framework for multi-robot gas source localization that achieves reliable calibration-free estimation by fusing rank-based local beliefs invariant to sensor heterogeneity, while simultaneously optimizing team coordination through informative region allocation and path planning.

Original authors: Wanting Jin, Marc Zoel Arias Mitjà, Alcherio Martinoli

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wanting Jin, Marc Zoel Arias Mitjà, Alcherio Martinoli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a team of small robots sent into a dark, cluttered building to find the source of a dangerous gas leak. Their mission is urgent: locate the leak quickly to prevent disaster. In the real world, these robots are rarely identical twins. They are often built with different sensors, perhaps bought at different times or from different manufacturers. Even if two sensors are designed to measure the same thing, they rarely agree on the exact number. One might read a concentration as high while another reads it as low, simply because of how they were made or how old they are. This lack of agreement makes it incredibly difficult for the robots to work together. If they try to combine their raw data, the conflicting numbers create confusion, and the team fails to find the source. This is the central challenge researchers at the École Polytechnique Fédérale de Lausanne set out to solve: how can a group of robots with mismatched, uncalibrated sensors collaborate to find a gas leak without needing to perfectly calibrate each one first?

The researchers developed a new way for these robots to share information that ignores the specific numbers each sensor reports. Instead of arguing over whether a gas reading is "five" or "ten," the robots focus on the order of the readings. They ask a simpler question: "Is the gas stronger here than it was a moment ago?" or "Is this spot smellier than that one?" By comparing the relative strength of the gas over time and space, rather than the absolute value, each robot can build a local map of where the leak might be. This method works even if one robot's sensor is extremely sensitive and another's is dull, because the ranking of the smells remains consistent regardless of the sensor's quirks. Once each robot has created its own local map based on these rankings, they share these maps with the team. The group then combines them, looking for the one location that all the robots agree is the most likely source. This process, known as fusing beliefs, allows the team to converge on a single, accurate answer even when their individual sensors are wildly different.

To make this search efficient, the team also designed a strategy to keep the robots from getting in each other's way. In many previous attempts, robots would stop at specific points to take a measurement, then move to the next point, often leading to redundant paths where multiple robots checked the same spot. The new approach allows the robots to gather data while they are moving, which is much faster. However, moving freely can lead to chaos if the robots cross paths too often. To solve this, the team divides the building into distinct zones, assigning each robot its own area to explore. They ensure that the zones cover the most promising areas first, based on the combined map, but also spread out to cover new ground. As the robots move through their assigned zones, they constantly update the team's shared map, which in turn helps the group decide where to go next. This balance between exploring new areas and focusing on the most likely spots allows the team to find the source without wasting time or energy.

The researchers tested this system in a highly realistic computer simulation that mimicked the complex airflow of an indoor environment with obstacles like walls and furniture. They used a team of three robots, each equipped with a sensor that behaved differently from the others, just as real-world sensors do. In one set of tests, the sensors were perfectly calibrated, and in another, they were left uncalibrated to reflect real-world imperfections. When the researchers compared their new method against a standard approach that simply added up all the raw numbers from the sensors, the difference was stark. The standard method failed completely when the sensors were uncalibrated, unable to find the source no matter how long the robots searched. In contrast, the new ranking-based method successfully located the gas source in almost every trial, regardless of whether the sensors were calibrated or not. The robots found the leak with high accuracy and did so without traveling excessive distances, proving that they could work together effectively even with imperfect hardware.

This work demonstrates that a group of robots does not need to be perfectly uniform to solve complex problems together. By shifting the focus from exact measurements to the relative patterns of those measurements, the team created a system that is robust to the inconsistencies of real-world sensors. The findings suggest that in emergency scenarios where time is critical and equipment may vary, relying on the collective intelligence of a diverse robot team is a viable and powerful strategy. The researchers note that while their results come from simulations, the principles they used are grounded in the physical behavior of gas and sensors, offering a promising path forward for deploying these systems in actual disaster zones where every second counts.

Technical Summary: Probabilistic Multi-Robot Gas Source Localization with Uncalibrated Sensors

Problem Statement
Estimating environmental states using Multi-Robot Systems (MRS) is critical for tasks like Gas Source Localization (GSL), particularly in emergency scenarios. However, practical deployment faces two significant hurdles:

Sensor Heterogeneity: Robots are often equipped with uncalibrated, heterogeneous sensors (e.g., Metal Oxide (MOX) gas sensors) that exhibit non-linear responses, varying sensitivities, and drift due to manufacturing differences and environmental factors (temperature, humidity, age). Standard probabilistic approaches that aggregate raw measurements assume homogeneous sensors or simple noise models, leading to unreliable fusion when sensor responses are inconsistent.
Coordination Efficiency: Effective multi-robot coordination is required to gather complementary information while avoiding redundant exploration. Existing strategies often rely on "stop-sense-go" paradigms or assume homogeneous sensing, which limits performance in dynamic, continuous sampling scenarios.

Methodology
The authors propose a distributed probabilistic framework that enables calibration-free estimation and coordinated active perception. The workflow operates in iterative cycles involving Source Term Estimation (STE), belief fusion, and path planning.

Rank-Based Source Term Estimation (STE):
- Instead of comparing absolute gas concentration values (which vary wildly between uncalibrated sensors), each robot extracts a rank-based gas feature.
- For a robot's measurement set $Z^m_k$ , the method computes the Empirical Distribution Function (EDF) to determine the normalized rank of each measurement relative to its own history. Similarly, ranks are computed for predicted concentrations derived from a Data-Driven Plume Model (DDPM) based on candidate source locations.
- The likelihood of a source hypothesis is calculated based on the alignment between the rank of measured data and the rank of predicted data. This approach is invariant to sensor scaling and non-linearities, assuming only that the sensor response is a monotonic transformation of the true concentration.
- Each robot independently updates a local belief map $b^m_k$ using Bayesian inference.
Distributed Belief Fusion:
- Robots exchange their local belief maps.
- A Product of Experts (PoE) formulation is used to fuse these beliefs into a global belief map $b_k$ . This formulation assumes conditional independence and reinforces hypotheses that are consistently supported across the team, suppressing outliers caused by individual sensor noise or bias.
Informative Region Allocation and Path Planning:
- To balance exploration and exploitation while minimizing redundancy, the environment is partitioned into $M$ non-overlapping, spatially connected regions.
- Region Allocation: Active cells (those with belief values above a threshold) are identified. The top- $M$ highest-probability cells serve as seeds for a multi-seed breadth-first search to assign distinct regions to each robot, ensuring balanced workload and connectivity.
- Informative Path Planning (IPP): Within its assigned region, each robot plans a trajectory toward an information-weighted centroid. This allows robots to continuously collect data while moving (sense-in-motion), shifting focus from broad exploration (when beliefs are diffuse) to focused exploitation (as beliefs concentrate).

Key Contributions

Calibration-Free Distributed Estimation: A framework allowing robots with uncalibrated, heterogeneous sensors to independently estimate local posteriors and fuse them into a consistent global estimate without requiring sensor calibration.
Rank-Based Feature Extraction: The use of rank-based features to handle sensor non-linearities and scaling differences, enabling reliable likelihood estimation where direct measurement aggregation fails.
Coordinated Exploration Strategy: An informative region allocation and path planning strategy that partitions the environment to reduce redundant observations while maintaining efficient information gathering during motion.
Extensive Validation: Evaluation in a high-fidelity Webots simulation using realistic MOX sensor models (incorporating non-linearity, noise, and slow dynamics) under both calibrated and uncalibrated conditions.

Results
The proposed "Belief Sharing" method was benchmarked against a "Measurement Aggregation" baseline (which aggregates raw data from all robots) across three indoor environments with varying obstacle configurations.

Uncalibrated Sensors: The baseline method failed to localize the source in uncalibrated scenarios, yielding large estimation errors and requiring excessive travel distances due to the mismatch between predicted and raw sensor values. In contrast, the proposed method successfully localized the source in most trials, demonstrating robustness to strong sensor heterogeneity.
Calibrated Sensors: Both methods achieved comparable accuracy and efficiency, confirming that the proposed method does not sacrifice performance when sensors are well-calibrated.
Efficiency: The proposed framework achieved a good balance between exploration and exploitation, localizing sources with moderate trajectory lengths.

Significance and Claims
The paper claims that this work demonstrates how calibration-free sensing representations can be effectively extended to distributed robotic systems. By decoupling the estimation process from absolute sensor values and relying on relative evolution (ranks), the system overcomes the limitations of sensor heterogeneity that typically plague multi-robot gas sensing. The authors position this as a step toward applying similar calibration-free representations to other distributed estimation tasks involving heterogeneous sensors. They acknowledge current limitations regarding communication scalability ( $O(M^2 N_e)$ ) and identify future work in reducing data exchange (e.g., incremental updates) and validating the framework in real-world experiments.

Technical Summary: Probabilistic Multi-Robot Gas Source Localization with Uncalibrated Sensors

More like this