Yes this can be done. I have read a lot about Q learning and SARSA(Lambda) used in this case too. You can check them too. Implementing Q and SARSA seem quite a bit easy to me rather than using GA. Check the papers of Prof Muhidul Khan from Hebut University.
I a bit doubt about the performance when executing some Genetic Algorithms at a sensor node because of its limited capacity. GAs is high cost, I believe. So, we should design centralized GAs for routing paths which is executed at a central node. And then sensor nodes just query for such routing information.
If you still would like to select paths at a sensor node, I think some simple mechanisms for learning about the good routing paths based on the exchange information between 1-hop or 2-hop neighbor nodes can be applied. Good luck.