Leveraging class abstraction for commonsense reinforcement learning via residual policy gradient methods