Source-linked AI summary
Lightweight Change Detection for Adaptive Collective Perception in Dynamic Environments
Thiemen Siemensma, Tiago Travassos, Jiacheng Zhang, Bahar Haghighat
TL;DR
Collective perception in dynamic environments remains underexplored, particularly when spatial correlations change alongside tile proportions. This paper adds lightweight change detection and online adaptive sampling to Bayesian collective decision-making. Simulations report competitive detection delays, more consistent precision across communication settings, and improved accuracy under strong spatial correlations.
Problem
Dynamic collective-perception studies have focused mainly on static environments or fill-ratio changes, while changing surface-pattern correlations remain underexplored despite affecting decision difficulty.
Source: §1 · p. 2Method
Results
The proposed approach achieves competitive detection delays and more consistent precision across communication settings, while adaptive sampling improves decision accuracy under strong spatial correlations.
Source: §6 · p. 12Takeaways & Limitations
In Webots simulations, lightweight change detection and adaptive sampling support dynamic collective perception under changing fill ratios and spatial correlations.
Source: §1 · p. 3, §6 · p. 12The correlation estimator may confuse environmental correlation with persistent-walk and boundary effects, and CUSUM can degrade when its reference absorbs post-change samples.
Source: §6 · p. 12
Abstract
from arXiv · showhide
Swarm robotics enables scalable inspection through collective perception, where robots infer environmental properties from distributed observations. A common benchmark considers robots exploring black-white tiled environments and determining the majority tile type. While most work studies static environments, dynamic environments remain underexplored. We study collective perception in dynamic environments where the tile proportion and the spatial correlation of surface patterns may change over time. To enable implementation on resource-constrained robots, we integrate a lightweight change detection mechanism into a Bayesian collective decision-making framework that resets robot beliefs when distribution shifts are detected. The proposed detector requires constant memory and constant per-sample computation, unlike two recent competitive approaches that store the full sample stream, making it suitable for memory-constrained robots. We further introduce a method that detects spatially correlated sampling and adapts sampling intervals online to mitigate correlation effects. Our simulation results in Webots show competitive change detection delays, more consistent precision across communication settings, and improved decision accuracy under changing spatial correlations.
1 Introduction
Collective perception asks robot swarms to infer the majority tile type from distributed observations, but dynamic environments and changing spatial correlations remain underexplored. The work targets lock-in and resource constraints with lightweight change detection and adaptive sampling.
1 Introduction
- Changing tile proportions and spatial correlations are underexplored in collective perception, despite their influence on decision difficulty.Existing dynamic-environment studies primarily consider fill-ratio changes.Source: §1 · p. 2
- Decision lock-in is an absorbing state that prevents robots from reversing decisions despite new observations.Prior mitigations include stubborn agents, discarding outdated information, and limiting communication.Source: §1 · p. 2
- Dynamic Bayes Bots reset decisions but are constrained by a belief cap and can reset slowly when the Beta posterior becomes inertia-heavy.DBBCPD adds PELT-based change detection and neighbor feedback, but stores the full sample stream, requiring O(n) memory and O(n^2) cumulative computation.Source: §1 · pp. 2–3
- The study considers fill-ratio shifts, clustering-driven correlation changes, and correlation changes caused by feature-size variation.The illustrated fill-ratio example shifts from 0.3 to 0.7.Source: Figure 1 · p. 3
- This work combines an O(1)-memory CUSUM fill-ratio detector with online spatial-correlation estimation and adaptive sampling, evaluated against DBB and DBBCPD.CUSUM-style control charts had not yet been applied to collective perception.Source: §1 · p. 3
2 Problem Statement
The problem is collective majority-tile estimation in a dynamic tiled environment, where fill ratio and spatial correlation can change. Robots estimate and adapt to spatial correlations while deciding the majority tile type over time.
2 Problem Statement
- Robots operate in a 1 m × 1 m black-and-white tiled environment and estimate the majority tile type over time.The fill ratio is the proportion of white tiles.Source: §2 · p. 3
- The environment may change through fill-ratio shifts, tile clustering, or feature-size variation that alters spatial correlation.Pre- and post-change values are denoted with subscripts 0 and 1.Source: §2 · p. 3
- Robots actively estimate and adapt to spatial correlations to enhance decision-making accuracy.Source: §2 · p. 3
3 Proposed Algorithm
The algorithm extends Bayesian collective perception with a CUSUM detector that resets beliefs after fill-ratio changes and a correlation detector that adjusts sampling distance. The extensions aim to retain lightweight change detection while reducing errors from spatially correlated samples.
Proposed extensions
- The proposed extensions comprise an O(1)-memory CUSUM detector and a mechanism for estimating and adapting to spatial correlations.The CUSUM chart accumulates deviations from the estimated fill ratio to detect changes in sample probability.Source: §3 · pp. 3–4
Sampling and decision-making
- Each robot updates a Beta posterior from local or received binary samples, uses its belief to decide the majority tile type, and retains the decision until contradiction warrants a change.The algorithm combines random-walk sampling with CUSUM resets and correlation-based step-size updates; it omits positive feedback because reinforced posterior updates can cause decision lock-in.Source: §3 · p. 4
Detecting fill-ratio changes
- CUSUM accumulates sample deviations from the posterior-mean fill-ratio estimate and resets the robot’s Beta model and decision after a detected change.Its two statistics accumulate upward and downward deviations; the sensitivity parameter sets negative drift under the no-change hypothesis.Source: §3 · pp. 4–5
- The CUSUM threshold is calibrated to a target false-alarm run length and adjusted online as the estimated sample count changes with sampling rate.The online adjustment keeps false-alarm probability independent of sample rate, although lower rates increase true-change detection delay.Source: §3 · p. 5
Detecting correlations
- The correlation extension compares empirical local flips with expected flips to estimate transitions per step and infer spatial sampling correlation.The inferred correlation is used to estimate tile size; that estimate is unavailable when the step exceeds tile size because the estimator saturates.Source: §3 · pp. 5–6
Adapting to correlations
- Robots adjust step size every m samples, increasing it when too few tiles are traversed and decreasing it when observed correlation is low.The increase is capped at twice the current step; around tile size, the update settles into bounded oscillation rather than a fixed point.Source: §3 · p. 6
4 Experimental Framework and Scenarios
The evaluation uses a physics-based Webots environment and a vibration-sensing robot model, comparing change-detection baselines across three scenarios. The scenarios assess fill-ratio changes, correlation estimation, and adaptive sampling, using accuracy and detection metrics.
Robot model and baselines
- The physics-based Webots simulations use a robot model with differential drive, collision-avoidance sensors, RF24 communication, and distance-based sampling via wheel encoders.Each simulated experiment concurrently runs DBB, DBBCPD, and CUSUM on the same samples for fair comparison.Source: §4 · p. 7
Experimental scenarios
- Experiments run for 1,200 seconds across 100 simulation runs per configuration, with environmental changes introduced halfway through.The duration matches the robot platform’s approximate battery lifetime.Source: §4 · p. 7
- The three scenarios test inverted fill-ratio changes, correlation estimation across tile sizes and step sizes, and fixed versus adaptive sampling under spatial correlation.The correlation-estimation scenario evaluates clustering and tile-size changes for false-positive detections while fill ratio remains constant.Source: §4 · pp. 7–8
Evaluation metrics
- Evaluation measures accuracy, detection delay, recall, and precision; missed changes count as false negatives and are excluded from delay statistics.Accuracy is one minus the absolute error between the robot’s belief and the true final decision.Source: §4 · p. 8
5 Results and Discussion
Across three scenarios, the results compare change detection under fill-ratio shifts, assess spatial-correlation estimation, and test adaptive sampling to reduce correlation-related accuracy loss.
5.1 Scenario 1: Detecting fill ratio Changes
- CUSUM matches DBBCPD under reduced communication and has superior beliefs under all-to-all communication, where DBBCPD’s false-positive resets lower precision.CUSUM’s precision is more consistent across communication settings, although neither detector consistently minimizes detection delay.Source: Figure 3 · p. 9, §5.1 · pp. 9–10, Table 1 · p. 9
5.2 Scenario 2: Detecting Correlations
- As spatial correlation increases, all methods lose accuracy; DBB has the smallest loss, while DBBCPD’s all-to-all false-positive resets lower its accuracy.CUSUM’s accuracy does not recover after the change in these results; the paper identifies a speed–accuracy trade-off for faster fill-ratio detection.Source: §5.2 · p. 10, Figure 4 · p. 11
- The tile-size estimator converges to the true size when d≤ℓ, but is biased toward d when d>ℓ; the selected convergence interval is m=29.For ℓ=0.2 m, boundary-reflection effects produce a slight positive bias.Source: Figure 5 · p. 11, §5.2 · p. 11
5.3 Scenario 3: Adapting to Correlations
- Adaptive step sizes improve decision accuracy by reducing false fill-ratio detections, particularly under strong spatial correlations.The comparison holds the fill ratio constant while spatial correlation varies.Source: Figure 6 · p. 12, §5.3 · p. 12
6 Conclusion and Future Work
The paper combines lightweight fill-ratio change detection with correlation estimation and adaptive sampling for dynamic collective perception. It reports competitive detection delays, more consistent precision across communication settings, and improved accuracy under strong spatial correlations, while identifying limits of the estimator and detector.
6 Conclusion and Future Work
- The proposed approach combines an O(1)-memory CUSUM fill-ratio detector, online sampling-correlation estimation, and adaptive distance-based sampling.Source: §6 · p. 12
- CUSUM has competitive detection delays and more consistent precision across communication settings, while adaptive sampling improves accuracy under strong spatial correlations.Source: §6 · p. 12
- The estimator cannot identify excessive decorrelation and may confound environmental correlation with persistent-walk or boundary effects; CUSUM can degrade as its reference absorbs post-change samples.Future work includes improved CUSUM references, simultaneous fill-ratio and correlation changes, and real-world validation with wheel-odometry slip and drift.Source: §6 · p. 12