Skip to main content

RBE MS Thesis Presentation: Nishan Srishankar | Distributed Bayesian Filtering for Collective Perception with Noisy and Byzantine Robots

robotics_banner.jpg

Various images of robots at Robotics Engineering WPI alt
WPI Robotics Engineering
Wednesday, May 05, 2021
10:00 am to 11:00 am

ROBOTICS ENGINEERING MASTERS THESIS PRESENTATION

 

Nishan Srishankar

Distributed Bayesian Filtering for Collective Perception with Noisy and Byzantine Robots

 

Wednesday, May 5, 2021

10:00 AM - 11:00 AM

Virtual | Zoom: https://us02web.zoom.us/j/2793898038?pwd=RjFXNWM0cmRTUS9kcGdOSmxXWEp3dz09

 

Abstract: A central problem in swarm robotics is collective perception. In practical settings, the robots must deal with uncertain information. Uncertainty stems from noisy sensors, occasional software bugs, or deliberate attacks that turn robots into defectors. With uncertain information, it is difficult for the robots to agree on the current state of the world that they are individually perceiving. Existing algorithms for collective perception typically do not consider uncertainty or, when they do, they focus on specific types of uncertainty, e.g., noise or malicious attacks, but not both at the same time. In this work, we focus on the problem of devising a collective perception algorithm that is resilient to both noise and defectors. We cast collective perception as a best-of-n decision-making problem in which robots must collectively agree on the current state of the world given a set of n possible states. Specifically, we assume that the robots are equipped with a ground sensor that can perceive black and white tiles. We provide robots with a set of possible ground patterns of black and white tiles. The robots must diffuse, communicate, and achieve consensus on the correct pattern. However, the ground sensor is imperfect in that it has a certain probability of returning the wrong color. In addition, certain robots are malicious: they communicate legal messages with deliberately incorrect information. We consider three idealized kinds of malicious robots, differing in the way they choose what to communicate:(i) Crazy robots individually pick a random message at every timestep with uniform probability;(ii) Stubborn robots individually choose a message at the beginning of the experiment and do not change it throughout;(iii) Zealot robots disseminate the same random message chosen at every timestep, Our approach is based on a Bayesian formulation of the collective decision process in which the robots combine individual and social information to produce a decision. Individual estimates are based on the output of the ground sensor. Social estimates are based on messages sent by neighbors. Every robot compares the received messages with the individual estimates and calculates a social estimate. The social estimate is a weighted sum of the messages that are sufficiently similar to the individual information.

To validate our approach, we ran experiments with swarms of {16,64,256} robots. We considered a fraction of these robots to be defectors in the set {0.0,0.25,0.50,0.75}. The sensor uncertainty is implemented as a probability to return the wrong color; we considered the probabilities {0.0,0.20,0.40}. We show that even in challenging environments, a majority of the swarm reaches consensus.

MS Thesis Advisor:

Professor Carlo Pinciroli, WPI

Committee Members:

Professor Berk Calli, WPI

Professor Xiangrui Zeng, WPI

DEPARTMENT(S):