LG AI Research 235

Complex real-world tasks have a compositional task structure in which multiple small subtasks are combined in a specific order. Existing methods have limitations in handling unseen tasks (unseen entities) where new objects appear.

To address this, the proposed PSGI (Parameterized Subtask Graph Inference) models tasks using a first-order logic approach. Instead of handling all subtasks individually, it infers structure using parameterized options (e.g., [Pickup x]).

This approach provides two key benefits:

• Compressibility of the graph: Since common structures are shared, data redundancy is reduced, enabling efficient inference with fewer samples.
• Zero-shot generalization: Tasks containing objects not seen during training (e.g., eggs, cabbage, etc.) can be instantly adapted to through the logical structure.

Experimental results showed that PSGI demonstrated higher efficiency and generalization performance than existing research in cooking, Minecraft-based mining, and AI2Thor simulation environments.