Byzantine Convergence in Robots Networks: The Price of Asynchrony

We study the convergence problem in fully asynchronous, uni-dimensional robot networks that are prone to Byzantine (i.e. malicious) failures. In these settings, oblivious anonymous robots with arbitrary initial positions are required to eventually converge to an a apriori unknown position despite a subset of them exhibiting Byzantine behavior. Our contribution is twofold. We propose a deterministic algorithm that solves the problem in the most generic settings: fully asynchronous robots that operate in the non-atomic CORDA model. Our algorithm provides convergence in 5f+1-sized networks where f is the upper bound on the number of Byzantine robots. Additionally, we prove that 5f+1 is a lower bound whenever robot scheduling is fully asynchronous. This constrasts with previous results in partially synchronous robots networks, where 3f+1 robots are necessary and sufficient.

💡 Research Summary

The paper tackles the convergence problem for a collection of oblivious, anonymous robots moving on a one‑dimensional line when a subset of them may behave arbitrarily (Byzantine faults) and the system operates under the fully asynchronous CORDA model. In this model, each robot repeatedly executes three phases—Look, Compute, and Move—but the phases are non‑atomic and the scheduler may interleave them arbitrarily, which makes it impossible for a robot to rely on a consistent snapshot of the whole system. The authors’ main contributions are twofold.

First, they present a deterministic algorithm that guarantees convergence for any initial configuration, provided the total number of robots n satisfies n ≥ 5f + 1, where f is an upper bound on the number of Byzantine robots. The algorithm works as follows. During a Look phase a robot obtains a (possibly outdated) multiset of positions of all robots that are currently visible. In the Compute phase it discards the f smallest and the f largest values from this multiset—these are the most extreme values that a Byzantine robot could inject. From the remaining positions it computes the arithmetic mean μ and then moves towards μ, but never more than half the distance to μ. Because the movement is limited, a robot that is still executing a previous move cannot be overtaken by a later move of another robot, preserving the monotonic shrinkage of the convex hull of the correct robots. The authors prove that after each complete round (i.e., after every correct robot has performed at least one Look‑Compute‑Move cycle) the interval that contains all correct robots shrinks by a factor strictly less than 1. Repeating this contraction infinitely many times forces the interval length to converge to zero, which implies that all correct robots converge to a common point. The proof carefully handles the worst‑case interleavings allowed by CORDA and shows that the contraction factor depends only on the ratio f/(5f + 1), guaranteeing convergence regardless of the scheduler’s choices.

Second, the paper establishes a matching lower bound: if n ≤ 5f, then no deterministic algorithm can guarantee convergence under the same adversarial conditions. The authors construct an adversarial schedule combined with Byzantine robots that continuously inject “fake” extreme positions on opposite sides of the correct robots. By carefully timing the activations, the scheduler ensures that each correct robot’s Look phase sees at most f extreme values, which are then removed by any reasonable filtering rule, leaving a set of positions that still spans a non‑shrinking interval. Consequently, the correct robots can be forced to oscillate forever between two disjoint regions, preventing convergence. This construction demonstrates that the factor 5f + 1 is not only sufficient but also necessary for convergence in the fully asynchronous setting.

The algorithm assumes only constant‑size memory (robots are oblivious) and requires each robot to store the list of observed positions temporarily for one Compute step. The computational cost per round is dominated by sorting the observed positions to discard the extremes, which is O(n log n) in the worst case, but can be reduced to linear time with selection algorithms. The movement rule is simple and does not depend on any global coordinate system or shared orientation, making the solution robust to the typical constraints of robot swarms.

Beyond the theoretical contribution, the authors discuss practical implications. In real‑world applications such as autonomous vehicle fleets, drone swarms, or distributed sensor networks, perfect synchrony is rarely achievable; communication delays, processing jitter, and environmental disturbances introduce asynchrony that closely matches the CORDA model. Moreover, malicious agents (e.g., compromised drones or spoofed GPS signals) can act as Byzantine nodes. The result that 5f + 1 robots are both necessary and sufficient gives system designers a clear quantitative guideline for sizing a resilient swarm: for every potentially compromised robot, at least five correct robots must be present to guarantee eventual agreement on a meeting point.

The paper concludes with several avenues for future work. Extending the algorithm to higher‑dimensional spaces, handling bounded movement speeds, or incorporating probabilistic schedulers could broaden its applicability. Another interesting direction is to investigate whether randomization can reduce the required number of robots below 5f + 1 in the asynchronous model, as it does in some consensus problems. Overall, the work closes a gap between synchronous and asynchronous Byzantine robot convergence, establishing the exact cost of asynchrony in terms of the minimal swarm size.