Scaling reinforcement learning through better representation and sample efficiency