This paper presents a novel distributed on-line evolutionary learning algorithm for swarm robotics that can cope with very limited hardware, as expected from using a swarm of low cost robots. The algorithm is able to deal with hardware constraints over the communication bandwidth by sharing only a limited amount of information, using a recombination operator inspired from bacterial conjugation. Using a classic foraging task, we show that the algorithm converges towards stable and efficient solutions even though, as expected, it converges slower when the bandwidth is limited. However, we also show that the proposed algorithm performs a trade-off between convergence speed and absolute performance that depends on the amount of bandwidth available. The recombination operator yields better performance if communication is limited, as recombination makes the most from the genetic material already present in the population. In other words, quality outweighs convergence speed if the bandwidth is limited.