Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Discovering Antagonists in Networks of Systems: Robot Deployment

Ingeborg Wenger, Peter Eberhard, Henrik Ebel

TL;DR

This paper addresses detecting antagonists in robot swarm surveillance by learning a context-conditioned action likelihood $p_a(a|s)$ from normal deployment data using a spline-based normalizing flow. It introduces a robust context embedding via a bidirectional LSTM and a refined action representation, and compares three detectors, with the mean-log-prob criterion delivering high sensitivity (>$80\%$) at a controlled false positive rate ($FPR\leq 5\%$). The method, tested on simulated scenarios and hardware with the HERA robots, outperforms the previous WengerEbelEberhard24 approach in both predictive performance and detection robustness, and is capable of operating without labeled anomaly data. The work demonstrates practical impact by enabling on-line identification and countermeasures against antagonists in swarm deployments, with code and data publicly available.

Abstract

A contextual anomaly detection method is proposed and applied to the physical motions of a robot swarm executing a coverage task. Using simulations of a swarm's normal behavior, a normalizing flow is trained to predict the likelihood of a robot motion within the current context of its environment. During application, the predicted likelihood of the observed motions is used by a detection criterion that categorizes a robot agent as normal or antagonistic. The proposed method is evaluated on five different strategies of antagonistic behavior. Importantly, only readily available simulated data of normal robot behavior is used for training such that the nature of the anomalies need not be known beforehand. The best detection criterion correctly categorizes at least 80% of each antagonistic type while maintaining a false positive rate of less than 5% for normal robot agents. Additionally, the method is validated in hardware experiments, yielding results similar to the simulated scenarios. Compared to the state-of-the-art approach, both the predictive performance of the normalizing flow and the robustness of the detection criterion are increased.

Discovering Antagonists in Networks of Systems: Robot Deployment

TL;DR

This paper addresses detecting antagonists in robot swarm surveillance by learning a context-conditioned action likelihood from normal deployment data using a spline-based normalizing flow. It introduces a robust context embedding via a bidirectional LSTM and a refined action representation, and compares three detectors, with the mean-log-prob criterion delivering high sensitivity (>) at a controlled false positive rate (). The method, tested on simulated scenarios and hardware with the HERA robots, outperforms the previous WengerEbelEberhard24 approach in both predictive performance and detection robustness, and is capable of operating without labeled anomaly data. The work demonstrates practical impact by enabling on-line identification and countermeasures against antagonists in swarm deployments, with code and data publicly available.

Abstract

A contextual anomaly detection method is proposed and applied to the physical motions of a robot swarm executing a coverage task. Using simulations of a swarm's normal behavior, a normalizing flow is trained to predict the likelihood of a robot motion within the current context of its environment. During application, the predicted likelihood of the observed motions is used by a detection criterion that categorizes a robot agent as normal or antagonistic. The proposed method is evaluated on five different strategies of antagonistic behavior. Importantly, only readily available simulated data of normal robot behavior is used for training such that the nature of the anomalies need not be known beforehand. The best detection criterion correctly categorizes at least 80% of each antagonistic type while maintaining a false positive rate of less than 5% for normal robot agents. Additionally, the method is validated in hardware experiments, yielding results similar to the simulated scenarios. Compared to the state-of-the-art approach, both the predictive performance of the normalizing flow and the robustness of the detection criterion are increased.

Paper Structure

This paper contains 18 sections, 10 equations, 10 figures, 1 table.

Table of Contents

  1. Introduction
  2. Antagonistic strategies during a surveillance task
  3. Detection of anomalous actions based on their probability of occurrence
  4. Finding a suitable representation for context and behavior
  5. Learning the likelihood of a robot action
  6. Criteria for categorizing an agent as antagonistic
  7. Setup of the deployment scenario for simulations and hardware experiments
  8. Preprocessing the state and action data
  9. Hardware and multibody model of the robot agents
  10. Training and tuning of the baseline and the proposed neural network
  11. Results
  12. Simulation results
  13. Hardware results
  14. Qualitative results
  15. Importance of the proposed methodological design choices
  16. ...and 3 more sections

Figures (10)

  • Figure 1: The Voronoi tessellation is used to determine the motion vectors of robots that exhibit normal behavior. The blue Voronoi cells show that the behavior of a robot might be invariant to the spatial translation of (local) objects, like neighboring robots, but not invariant to rotation within the inertial frame of reference.
  • Figure 2: A deployment scenario visualizing normal swarm behavior and the behavior of the different antagonistic strategies. The antagonist's region of interest is either marked with a circle in the corresponding color or located at the center of the illustrated density. The true motion of the spoofing robot toward its target is indicated by the dashed arrow.
  • Figure 3: Omnidirectional mobile robot (left) used in the hardware robot swarm (right) Ebel21. The blue lines indicate the deployment area.
  • Figure 4: Specificity, sensitivity and precision results for the baseline and the proposed network on the naive detection criterion, binomial detection criterion, and mean detection criterion. Higher values indicate a better performance. The allowed maximal false positive rate of normal agents is set to 5%.
  • Figure 5: Results for the mean criterion using an allowed maximal $FPR_{\max}$ of 1%.
  • ...and 5 more figures