LG AI Research 206

Existing reinforcement learning approaches have the limitation of low data efficiency, as they learn only a single policy to solve a specific problem. To address this, research on a Single General-purpose Policy capable of controlling various robots simultaneously is drawing attention.

The core challenge in controlling multi-joint robots is that each robot has a different structure and size. Existing Graph Neural Network (GNN) approaches suffer from an Over-smoothing problem, where information between nodes gets mixed during the message passing process, while Transformer approaches have the limitation of not sufficiently reflecting the robot's physical morphological information.

The model proposed to overcome this is SWAT(Structure-aWAre Transformer Policy). SWAT combines a Structural Embedding that reflects the robot's morphology with a Transformer model.

The structural embedding consists of the following:

• Positional Embedding: Expresses node positions on the graph through a tree traversal method
• Relational Embedding: Expresses relationships between nodes through the Laplacian matrix, shortest distance, personalized PageRank, and others

The SWAT model goes through three stages—'encoding-message passing-decoding'—to determine the optimal action that reflects both the robot's sensor information and structural features.