Long-horizon language-conditioned manipulation