GraspGPT (IEEE RA-L 2023)

Task-oriented grasping in real-world settings depends on how an object should be grasped (e.g., grasp a knife by its handle to cut fruit; grasp the blade to hand it to a user).

Existing approaches often rely on a pre-built knowledge base, which limits generalization to unseen objects and tasks. We propose a Transformer-based model that leverages LLM semantic knowledge for better generalization.

• Paper: https://arxiv.org/pdf/2307.13204
• Project/Video: https://sites.google.com/view/graspgpt/