how are you planning on getting the robots to learn a base policy from scratch? seems hard without a base model