Robot swarms coordinate through local communication links, but some links may be compromised and send arbitrary malicious data. A new paper from researchers including Thomas Kaminsky and Stephanie Gil proposes a wrapper that sits around robot communication and decides which incoming links can be safely used. The wrapper relies on trust values produced by an anomaly detector on each robot, and only requires that detector to be better than random chance.
The algorithm uses Safe Anytime-Valid Inference to test the stream of trust values. It offers two statistical guarantees at once: with a user-chosen high probability, it excludes every compromised link for all time, and it also quickly learns to trust legitimate links. According to the authors, no previous work provides both guarantees simultaneously without restrictive assumptions about connectivity or the number of compromised links.
The authors tested the wrapper on hardware for two tasks: neural implicit mapping with three LoCoBots and heading consensus with eight GoPiGos. They also ran simulations using trust values sampled from hardware, showing the approach scales to swarms of hundreds of robots. The paper is 44 pages with 10 figures and 6 tables.