I guess I hacked my own primitives reward function. Now I need to rediscover what this is all about. My naive approach is inverse reinforcement learning.