A Benchmark for Gas Source Localization toward Rapid Leak Removal in Cleanrooms

iCIR Lab · Department of Mechatronics Engineering

Chungnam National University

Demo Video

A complete trial in an open area of the cleanroom with moderate source strength and spread, from LRS patrol through HRS exploration to the final source-position decision.

Demonstration method: M6 · 4× speed


SAME ENVIRONMENT · DIFFERENT CASES

Special-case videos

Eight search conditions in the cleanroom: obstacles, initial LRS signal strength, source distance at the HRS transition, and sources inside inaccessible equipment or beyond a wall. Each clip follows exploration through the final source-position decision.

8 cases · 4× speed · Looping playback

CASE 014× speed

Open-area search

HRS exploration in an open part of the cleanroom, from the first target to the final source-position decision.

CASE 024× speed

Search near an obstacle

HRS exploration around equipment that restricts access, showing detours and the final source-position decision.

CASE 034× speed

Search after a weak LRS signal

Low but detectable LRS concentration readings lead into HRS exploration and the final source-position decision.

CASE 044× speed

Search after a strong LRS signal

High LRS concentration readings lead into HRS exploration and the final source-position decision.

CASE 054× speed

Search for a nearby source

With moderate LRS readings, HRS begins a short path distance from the source and continues to the final position decision.

CASE 064× speed

Search for a distant source

With moderate LRS readings, HRS begins a long path distance from the source and continues to the final position decision.

CASE 074× speed

Source inside inaccessible equipment

The gas source is inside equipment that the robot cannot enter. The clip shows the search under this access constraint.

CASE 084× speed

Source beyond a wall

The gas source is on the other side of a wall. The clip shows exploration with the wall separating the robot from the source.

Demonstration method: M6 · All eight cases

Abstract

When a hazardous gas leak occurs in a structured indoor environment such as a cleanroom, it is important to reach and remove the gas source rapidly while minimizing the need for personnel to enter the affected area. Cleanrooms with stringent cleanliness requirements, such as semiconductor front-end fabrication facilities, use vertical unidirectional airflow from the ceiling to a perforated floor. In these environments, the horizontal airflow component is negligible, leaving no wind-direction cues that indicate the direction of the gas source.

Most research on mobile robot-based gas source localization (GSL) relies on plumes generated by horizontal airflow and wind-direction measurements as key cues, while established benchmarks focus on warehouses, residential environments, and wind tunnels. To the best of our knowledge, no GSL evaluation has been reported for the vertical unidirectional airflow conditions found in cleanrooms. Moreover, existing studies typically assume that the search begins without prior information and focus on localization accuracy or concentration mapping. Consequently, how quickly a robot can reach a gas source using limited concentration measurements collected during routine patrols remains insufficiently explored.

This paper proposes a benchmark for a single gas source that evaluates the process from anomaly detection during a robot's patrol to the final source-position decision in the absence of wind-direction cues. In regions away from obstacles, gas concentration on the horizontal plane at sensor height is modeled as an isotropic Gaussian, since there is no horizontal airflow to skew the distribution. The robot measures concentrations at predefined locations during a low-resolution survey (LRS). Upon detecting an anomalous concentration, it switches to a high-resolution search (HRS), using a Gaussian Markov random field (GMRF) to estimate the concentration field and locate the source. Evaluation measures the time from the transition to HRS until the final source-position decision, together with the source localization error. Methods are compared by a Score that averages, over trials, the ratio of ideal travel time to HRS decision time, discounted by the localization error expressed in GMRF cells. HRS decision time, localization error, and the time to reach the final estimated position are reported as detailed metrics.

We randomly generated gas source positions, strengths, and distribution widths in an indoor environment with obstacles simulated using ROS 2, Gazebo, and Nav2. We compared six methods combining three search starting-point selection rules (the patrol location with the highest measured concentration, the concentration-weighted centroid of patrol locations, and the cell with the highest estimated mean) with two subsequent search strategies (selection of the cell with the highest estimated mean and spiral search). All methods were evaluated on the same 100 gas sources generated with fixed random seeds, and the 91 sources on which all six methods finished normally were used for comparison.

Combining the concentration-weighted centroid start with spiral search achieved the highest Score (0.311), followed by the highest-concentration patrol start with spiral search (0.277). For every starting-point rule, spiral search reached the source-position decision faster than selecting the cell with the highest estimated mean. Mean localization error was below 0.07 m for all six methods.

System overview

One benchmark trial: from a patrol pass to a final source-position decision.

LRS · One patrol pass

Measure concentration and update GMRF at each patrol stop

PATROL
Anomaly detected · complete the patrol
Search termination condition met

Decide source position

Record the estimate and decision time

TRIAL COMPLETE

Evaluate

Time · Error · Score

The operational system patrols continuously. This benchmark evaluates one trial at a time; the loop shown here is the repeated measurement and estimation within HRS.

OBSERVATION PROTOCOL

Measurement

Where concentration observations come from, how they are collected, and what is recorded.

  1. 01

    Define the position

    Use the robot center in the map frame as the horizontal sampling reference.

  2. 02

    Arrive and sample

    Collect concentration samples during the configured dwell interval at each target.

  3. 03

    Average and record

    Store the sample mean with the visit position, phase, time and sample count.

FRONT Physical inlet Sampling reference Robot center (x, y)
Conceptual top view; dimensions and mounting positions are not to scale.

A centered sampling reference

The physical inlet is offset from the robot center. For this experiment, the horizontal sampling and recording reference is placed at the robot center (base_link). The body’s inlet offset is not applied to the sampled coordinates.

Concentration is sampled from a static 2D Gaussian field at the virtual sensor’s (x, y). Its value is relative and bounded between 0 and 1; the computation does not use height z.

Position reference
Robot center · map frame
Nominal update rate
20 Hz
LRS dwell
1 s per visit
HRS dwell
2 s per visit

These are the current model/configuration defaults, not fixed requirements. Use the actual settings and sample counts for each experiment.

ONE VISIT → ONE OBSERVATION

Averaging the samples

The arithmetic mean of the samples collected during the dwell interval becomes the concentration observation used to update the GMRF.

c̄_j = (1 / n_j) Σ c_j,k
j
Measurement visit index.
c_j,k
The k-th concentration sample received during visit j (relative value, 0–1).
n_j
Actual number of samples received during that visit.
c̄_j
Mean concentration stored as one observation.
Recorded per visit

Position at the start of sampling · mean concentration · actual sample count · LRS/HRS phase · measurement event time. The recorded position is mapped to a GMRF cell for the field update.

FROM OBSERVATIONS TO PERFORMANCE

Evaluation Metrics

Compare decision time and localization error together. Use an ideal travel-time reference to express the decision-time ratio.

01 / TIME · s

Decision time

How quickly does the robot decide where the source is? Lower is better, alongside error.

T_HRS = t_decide − t_switch
T_HRS
HRS decision duration (s).
t_switch
Simulation time when the robot switches to HRS (s).
t_decide
Simulation time when the robot commits to the final source coordinates (s).

All methods use the same LRS procedure. The primary time comparison starts at the HRS transition and ends at the source-position decision.

LRS and total-time definitions
T_LRS = t_switch − t_start; T_total = T_LRS + T_HRS
t_start
Patrol start time (simulation seconds).
T_LRS
Patrol-start to HRS-transition interval, including any transition gap (s).
T_total
Patrol-start to final-decision duration (s); excludes subsequent approach.
02 / ACCURACY · m

Localization error

How close is the final estimated source position to the true source? Lower is better.

E = ‖p̂ − p_src‖₂ = √[(x̂ − x_src)² + (ŷ − y_src)²]
E
Straight-line localization error (m).
p̂ = (x̂, ŷ)
Final estimated source coordinates in the map frame (m).
p_src
True source coordinates (x_src, y_src), used as ground truth for evaluation (m).
‖·‖₂
Euclidean norm: straight-line distance between the two positions.

If the true source lies in a non-navigable area (inside equipment, walls or safety margins), the robot cannot reach it, so the nearest reachable point replaces p_src as the reference.

How the final estimate is selected
j* = arg max c̄_j over HRS visits; p̂ = q(p_j*)
M_HRS
Indices of valid HRS measurement visits.
j*, c̄_j
Visit with the highest mean concentration, and the mean defined in Measurement.
p_j, q(·)
Recorded visit position (m); q maps it to its GMRF cell center.

The current branch freezes the cell associated with the highest recorded HRS mean concentration. Equal means retain the earlier record. GMRF estimates support exploration; the final logged position follows this measured-maximum rule.

03 / REFERENCE RATIO

Decision-time ratio

How fast is the decision relative to ideal travel time? Compare with position error.

η = T* / T_HRS; T* = L* / v_max
η
Dimensionless ideal-travel-time / decision-time ratio.
T*
Ideal time to traverse the reference shortest path at the configured maximum speed (s).
L*
Shortest feasible path length from the robot’s HRS-transition position to the true source (m).
v_max
Maximum speed configured for the robot in the experiment (m/s). Use that experiment’s setting.
T_HRS
Decision duration defined above (s); positive for this ratio.

T* is an evaluation reference. The ratio may exceed 1 when the robot decides the source location before an ideal traversal would finish. Interpret it together with localization error.

Reference-path conditions

Compute L* with the same map and navigation constraints across methods. T* omits acceleration, turning and stops. If the true source is unreachable, define a reachable reference endpoint separately before computing the ratio.

04 / AGGREGATE SCORE

Score

Average the trial-wise time ratios, each discounted by that trial’s localization error, to compare methods on the same trial set. Higher is better.

Score = (1/N) Σ η_i / (1 + E_i/d) = (1/N) Σ (T*_i / T_HRS,i) · d / (d + E_i)
N
Number of trials included in the comparison for each method. Use the same trial set across methods.
i
Trial index, from 1 to N.
η_i
Time ratio of trial i. Each trial is weighted equally.
T*_i
Ideal reference travel time for trial i (s).
T_HRS,i
HRS decision duration for trial i (s).
E_i
Localization error of trial i, as defined in 02 (m).
d
GMRF cell size, 0.5 m. Final estimates are cell centers, so this is the resolution of the estimate.

E_i/d is the error in GMRF cells: zero error (E_i = 0) keeps η_i, a one-cell error halves it, and a two-cell error leaves one third. No success threshold is applied. Per-trial values are averaged, rather than dividing the average reference time by the average decision time.

Supplementary measure: time to estimated source
T_arrival = t_arrival − t_switch
t_arrival
Robot’s arrival time at the final estimated source position under the stated navigation criterion (simulation s).
t_switch
HRS transition time (simulation s).
T_arrival
Elapsed time from HRS transition through decision and approach to arrival (s).

This supplementary measure evaluates physical approach. Report it with localization error; arrival at an estimated position does not establish arrival at the true source.

The next section compares M1–M6 by Score on 91 shared seeds, with localization error and decision time in the details. Decision time is measured to the decision event, separately from arrival.

BENCHMARK RESULTS

Evaluation results

Comparing how quickly and accurately M1–M6 found the gas source, using Score.

91 shared seeds · 546 runs · Only seeds on which all six methods finished normally are included; each method ran independently under the same source conditions.

  • Top ScoreM6

    Score 0.311

Score

A single number for how quickly and accurately the gas source was found. Finding it faster and with a smaller error raises the score.

  1. M60.311
  2. M20.277
  3. M40.273
  4. M50.267
  5. M10.244
  6. M30.239
Mean · higher is better · sorted by Score

T* is a reference value computed from a reconstructed shortest path from the first logged HRS plan position (v_max = 0.18 m/s).

Show detailed results by method

Source localization error

Mean distance between the estimated position and the source (reachable reference).

  1. M60.056
  2. M20.022
  3. M40.038
  4. M50.064
  5. M10.023
  6. M30.051
Mean · meters (m) · lower is better · sorted by Score

Search decision time

Elapsed simulation time from the transition to HRS to the final source-position decision.

  1. M6199.8
  2. M2248.1
  3. M4245.7
  4. M5225.0
  5. M1269.7
  6. M3269.0
Bar: median · Line: Q1–Q3 · seconds (s) · lower is better

Summary by method

MethodStart ruleSearchScoreDecision time (s)
Median [Q1–Q3]
Decision time (s)
Mean ± SD
Mean error (m)0 m error
rate
M6Weighted centroidSpiral0.311199.8 [162.4–250.9]209.9 ± 62.30.05690.1%
M2Patrol maxSpiral0.277248.1 [212.1–278.3]243.0 ± 58.30.02295.6%
M4Estimated maxSpiral0.273245.7 [204.6–281.1]241.2 ± 59.90.03892.3%
M5Weighted centroidEstimated-max cell0.267225.0 [181.4–288.1]242.0 ± 83.50.06489.0%
M1Patrol maxEstimated-max cell0.244269.7 [236.3–315.3]279.9 ± 66.70.02396.7%
M3Estimated maxEstimated-max cell0.239269.0 [235.3–319.1]277.1 ± 61.10.05193.4%

Sorted by Score; bold values are the best in each column.

Data and selection criteria

Cohort: 91 seeds on which M1–M6 all finished with classification = normal. Excluded seeds (9): 103, 107, 114, 124, 138, 140, 145, 175, 182.

Supplementary: median time from the HRS transition to arrival at the estimated position is 207.8–280.2 s. LRS patrol time is excluded from all time metrics.

Download per-method summary (CSV)