• Proposes a novel task of instruction-driven 3D object affordance segmentation from point clouds, bridging natural language and spatial manipulation reasoning.
• Introduces the Instruction-Affordance Dataset (IAD) with 7,190 instances across 20 object categories and 624 manipulation instructions, including seen and unseen splits to test generalization.
• Designs the IDAS network that integrates language instructions with point cloud features layer-by-layer, directly outputting task-relevant manipulation regions.
• Demonstrates superior performance over existing methods under both seen and unseen instructions, showing strong generalization to novel commands and unknown affordances.