Task-Oriented Deep Reinforcement Learning for Robotic Skill Acquisition and Control.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Deep reinforcement learning (RL) and imitation learning (IL) are key for autonomous robotic control.
• Current methods demand extensive training interactions, limiting real-world robot deployment.

Purpose of the Study:

• To develop an efficient model-free, off-policy actor-critic algorithm for robotic skill acquisition and continuous control.
• To reduce the sample complexity and improve the performance of training autonomous robots.

Main Methods:

• Fusing task reward with a task-oriented guiding reward derived from few, imperfect expert demonstrations.
• Utilizing an actor-critic framework for intentional exploration and effective experience exploitation.

Main Results:

• Achieved 2-10 times lower sample complexity compared to state-of-the-art deep RL algorithms on robotic locomotion tasks.
• Surpassed expert performance in robotic locomotion tasks.
• Demonstrated significant improvements in sampling efficiency and asymptotic performance on sparse and delayed reward tasks.

Conclusions:

• The proposed algorithm substantially advances autonomous skill acquisition for real-world robots.
• The method enhances both learning speed and final performance, particularly in challenging reward scenarios.