Efficient State Representation Learning for Dynamic Robotic Scenarios