Understanding Dexterous Hand Grasping in Embodied Intelligence

Dexterous hand grasping is a key task in embodied intelligence; its high degrees of freedom make it one of the most difficult challenges.

1. Ego‑centric (robot‑centric) approaches

Ego‑centric methods directly map the robot’s observations and state to action commands, enabling fast inference. However, they demand extensive information processing and struggle to generalize across different hand configurations.

Representative work: UniDexGrasp (https://arxiv.org/pdf/2303.00938)

2. Object‑centric approaches

Object‑centric methods estimate an intermediate representation (e.g., a contact heatmap or contact points) from the observed object point cloud, then solve inverse kinematics to obtain the end‑effector pose. This optimization integrates object information into the grasp decision but incurs slower inference.

Representative work: GraspTTA (https://arxiv.org/pdf/2104.03304)

3. Interaction‑centric approaches

Interaction‑centric methods model the interaction between robot and object, constructing intermediate representations such as distance matrices, IBS surfaces, or CGR. From these, the current hand pose is estimated, offering both fast inference and better generalization to various hands.

Representative work: DRO_Grasp (https://arxiv.org/pdf/2410.01702)