Video Generation with Learned Action Prior