As robot swarms move to real-world deployment, safety will be a key factor in improving adoption and trust [45]. Robot swarms are composed of many robots: during real-world operation each individual may be susceptible to failure resulting in potentially degraded performance of the swarm overall. A necessary component for safety is then the ability to detect and mitigate faults in the swarm. In this paper, we present a novel approach to learning dynamic fault mitigation via neuroevolution, where mitigation actions are implemented by both faulty and non-faulty robots in a collective transport scenario. In particular, there is no explicit fault detection step and the evolved “mitigation module” maps between a set of locally observed metrics as input and mitigation actions as output. Our approach is able to learn effective mitigation for six types of fault independently. We show that by allowing robots of any state to freely apply actions, “loosely-coordinated” mitigation emerges improving on the baseline where no mitigation is applied.