Learning Robotic Manipulation Tasks via Task Progress Based Gaussian Reward and Loss Adjusted Exploration

Multi-step manipulation tasks in unstructured environments are extremely challenging for a robot to learn. Such interlace high-level reasoning that consists of the expected states can be attained achieve an overall task and low-level decides what actions will yield these states. We propose model-free deep reinforcement learning method learn multi-step tasks. introduce Robotic Manipulation Network (RoManNet) 1 , which is vision-based model architecture, action-value functions predict action candidates. define Task Progress based Gaussian (TPG) reward function computes on lead successful motion primitives progress towards goal. To balance ratio exploration/exploitation, we Loss Adjusted Exploration (LAE) policy determines from candidates according Boltzmann distribution loss estimates. demonstrate effectiveness our approach by training RoManNet several robotic both simulation real-world. Experimental results show outperforms existing methods achieves state-of-the-art performance terms success rate efficiency. The ablation studies TPG LAE especially beneficial like multiple block stacking.