Collision Avoidance and Liveness of Multi-agent Systems with CBF-based Controllers

TL;DR

This paper evaluates collision avoidance strategies for multi-agent systems using Control Barrier Functions, comparing decentralized and centralized policies, and introduces a new policy that balances safety and liveness effectively.

Contribution

It introduces a new decentralized collision avoidance policy (CCS2) and compares its performance with existing policies and a centralized benchmark, highlighting trade-offs between safety and liveness.

Findings

Decentralized policies lack liveness compared to centralized ones.

The PCCA policy performs as well as the centralized controller.

Feasibility is established for all considered policies.

Abstract

In this paper we consider multi-agent navigation with collision avoidance using Control Barrier Functions (CBF). In the case of non-communicating agents, we consider trade-offs between level of safety guarantee and liveness - the ability to reach destination in short time without large detours or gridlock. We compare several CBF-based driving policies against the benchmark established by the Centralized controller that requires communication. One of the policies (CCS2) being compared is new and straddles the space between policies with only local control available and a more complex Predictor-Corrector for Collision Avoidance (PCCA) policy that adjusts local copies of everyone's control actions based on observed behavior. The paper establishes feasibility for the Centralized, PCCA and CCS2 policies. Monte Carlo simulations show that decentralized, host-only control policies lack liveness compared to the ones that use all the control inputs in calculations and that the PCCA policy performs equally well as the Centralized, even though it is decentralized.