Goal-Driven Transformer for Robot Behavior Learning from Play Data

Robot behavior learning has emerged as a crucial field, allowing robots to adapt and improve their actions based on experiential knowledge rather than being solely reliant on predefined instructions. However, the effectiveness of such learning is often hindered by the limitations of offline reinforcement learning, which relies on pre-defined reward labels, and traditional imitation learning, which depends on high-quality expert demonstrations. To address these challenges, in this paper, we propose a novel Goal-Driven Transformer (GDT) for robotic behavior learning from play data. The core module of the GDT is the inclusion of the Goal-Driven Attention Block (GDAB) that utilizes attention mechanisms to concentrate the model’s focus on particular objectives, enabling the GDT to selectively focus on critical parts of the observation data to perform behavioral learning for specific goals. Moreover, we employ the Standard Attention Block (SAB) to ensure that this goal-directed learning occurs with a comprehensive understanding of the environment and the sequence of actions required. Experimental validation of the proposed GDT framework is conducted in two simulated environments: Block-pushing and Franka Kitchen. The results demonstrate that the GDT framework has achieved state-of-the-art performance in the realm of robot behavior learning from play data. Videos are available at: https://gdt-bl.github.io/.