This paper proposes an interesting method that learns multiple composable policies that can produce a continuous spectrum of diverse actions, where primitives can be transferred to new tasks. The paper is well written, and the experimental results are convincing. The ablation study would strengthen the paper.