BayesCPF: Enabling Collective Perception in Robot Swarms with Degrading Sensors

TL;DR

BayesCPF tackles collective perception with robots whose sensing degrades over time by jointly estimating the environment fill ratio $f$ and individual sensor accuracy $b[k]$ in a decentralized manner. It integrates the Minimalistic Collective Perception (MCP) algorithm for fill-ratio estimation with an Extended Kalman Filter (EKF) to track sensor degradation, coupling the two modules recursively so that $b[k]$ informs $f$ estimates and vice versa. The architecture comprises four modules—Fill Ratio Estimation (FRE), Sensor Accuracy Estimation (SAE) via EKF, Constraint Compliance (CC), and Observation Quantity Adjustment (OQA)—to maintain robust performance under time-varying degradation, while ensuring computational efficiency ($O(1)$ per-robot). Empirical results from simulations and physical experiments show BayesCPF approaches the performance of a system with known sensor accuracy in the transient phase and remains competitive under non-catastrophic degradation, though equilibrium-phase performance can be limited in highly ambiguous or severely degraded settings. The work demonstrates a practical, scalable approach to maintaining reliable collective perception in swarms of resource-constrained robots and points to future extensions for non-monotonic degradation and dynamic environmental changes.

Abstract

The collective perception problem -- where a group of robots perceives its surroundings and comes to a consensus on an environmental state -- is a fundamental problem in swarm robotics. Past works studying collective perception use either an entire robot swarm with perfect sensing or a swarm with only a handful of malfunctioning members. A related study proposed an algorithm that does account for an entire swarm of unreliable robots but assumes that the sensor faults are known and remain constant over time. To that end, we build on that study by proposing the Bayes Collective Perception Filter (BayesCPF) that enables robots with continuously degrading sensors to accurately estimate the fill ratio -- the rate at which an environmental feature occurs. Our main contribution is the Extended Kalman Filter within the BayesCPF, which helps swarm robots calibrate for their time-varying sensor degradation. We validate our method across different degradation models, initial conditions, and environments in simulated and physical experiments. Our findings show that, regardless of degradation model assumptions, fill ratio estimation using the BayesCPF is competitive to the case if the true sensor accuracy is known, especially when assumptions regarding the model and initial sensor accuracy levels are preserved.

Figures (14)

• Figure 1: An example application where swarm robots are deployed to collectively assess the bleaching coverage of coral reefs. Image classification models (that identify the presence of bleaching) onboard the robots may be less accurate due to suboptimal lighting conditions, as experienced by the robot on the left. Accordingly, it makes more detection mistakes due to a worse detection accuracy $b$.
• Figure 2: Environment setup for collective perception using a robot swarm. The orange circles indicate moving robots while the green lines indicate communication links. The red dotted line around a robot denotes the communication neighborhood of that robot.
• Figure 3: The flow of information in the Bayes Collective Perception Filter with the focus on robot $i$. Robot $i$ shares its local estimate (of the environment fill ratio $f$) $\hat{x}_i[k]$ with its neighbor, robot $j$, as indicated by pink lines. It also receives robot $j$'s local estimate, $\hat{x}_j[k]$ (green lines). Internally, robot $i$ uses the received $\hat{x}_j[k]$ to compute a weighted moving average informed estimate$x_i^\prime[k]$ that it uses to determine the (constrained) assumed sensor accuracy$\hat{b}_i[k]$. $\hat{b}_i[k]$ in turn influences the observations that robot $i$ use in estimating the fill ratio and sensor accuracy.
• Figure 4: An example of true sensor accuracy degradation behavior as governed by the Wiener process (\ref{['eq:wiener_process']}) with $\eta = -1 \times 10^{-5}$ and $\gamma = 1 \times 10^{-4}$, each curve representing one of 15 robots. The lowest possible accuracy level is fixed at 0.5 as this is the point where the sensor returns a completely random observation.
• Figure 5: Informed estimation errors when using the BayesCPF on robots with catastrophic sensor degradation ($b[k] \rightarrow 0.5$); lower values indicate better performance. Each box (and its outliers) contains data points from $M = 30$ trials. The initial true accuracy $b[0]$ is either $1.0$ (left column, secondary $x$-axis) or $0.8$ (right column, secondary $x$-axis). $x[k]$ estimation is better in situations where the modeling assumptions are preserved, i.e., in the transient phase where the state transition model follows true sensor degradation behavior.