Task-Oriented Grasping (TOG) requires a robot to pick up an object in a way that suits a downstream goal. For example, a robot asked to slice bread should grasp a knife by the handle, while a robot asked to hand the knife to someone might need a different hold. Choosing the right grasp is difficult because the same object affords different actions depending on the task.
TGrasp6D addresses this by combining task-informed vision-language priors with a pose refinement stage. The vision-language component helps the robot reason about which part of an object is relevant for a given task, and the refinement stage adjusts the grasp pose accordingly. Because the priors come from pretrained vision-language knowledge, the method can generalize to new tasks and objects without additional training.
This approach is notable for targeting full 6DoF grasping rather than simpler planar grasps, and for doing so in a zero-shot setting. The work suggests that grounding robotic manipulation in language and visual semantics could make task-aware grasping more flexible in real-world settings.