Graph Attention-Based Robotic Policy for Efficient Robot Learning

Increasing interest in advanced robotics has spurred a renewed concentration on developing efficient robot learning algorithms. A large portion of reinforcement learning research has focused on learning complex behaviors by leveraging the prior knowledge of human experts. This prior knowledge is typically summarized by human experts for specific robotic tasks, resulting in policies that are highly specialized and not easily transferable to more complex tasks. To resolve these limitations, we propose a model called Graph Attention-based Robotic Policy (GARP) that enhances robot learning efficiency by adopting the modified graph attention network. The modified graph attention network is specially designed to autonomously learn the relationships among joints, thereby deriving more efficient policies. Extensive experiments in continuous robotic environment, including the centipede task, characterizes that our method’s effectiveness to learn optimal policies for tackling robotic tasks, demonstrating extraordinary improvements over other competitive methods.