Long-horizon language-conditioned policy learning