Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives

Neural Information Processing Systems 

These parameterized primitives are expressive, simple to implement, enable efficient exploration and can be transferred across robots, tasks and environments.