Synthesize Policies for Transfer and Adaptation across Tasks and Environments

We further propose new training methods to disentangle the embeddings, making them both distinctive signatures of the environments and tasks and effective building blocks for composing the policies.