Winning back the CUP for distributed POMDPs: planning over continuous belief spaces