It's not clear to me how this is interestingly different from model-based RL, where you learn the state function and reward function, and then use various types of simulation to learn a value function. I guess I'll have to read more than the abstract...

Section 3.2 shows the successor representation (SR) definition. If I'm reading it correctly the SR might also be described as the discounted stationary distribution over states.

I haven't seen SR before in the RL literature, but the paper argues that this representation is useful for sub-goal identification. I guess I'll have to read more than the abstract as well :)