"Semantic" is the upgrade that turns a grasping demo into a useful robot. Google's grant US11717959B2 ("Machine learning methods and apparatus for semantic robotic grasping," issued August 8, 2023; inventors Eric Jang, Sudheendra Vijayanarasimhan, Peter Pastor Sampedro, Julian Ibarz, Sergey Levine) claims grasping that doesn't just find any feasible grip but grasps the object you asked for — the apple, not whatever's nearest. The grasp is conditioned on object identity.
The mechanism couples object recognition with grasp prediction in a learned model: the network is told (or infers) which object to retrieve and produces a grasp for that object specifically. B25J 9/163 (program-controlled manipulation) marks the action; G06N 3/008, 3/045, and 3/08 mark the neural control and training. The semantic conditioning is the delta over generic grasp-learning — and it's what a real task ("bring me the cup") requires.
For the manipulation beat, this is the same author lineage as Google's 2021 deep-grasping grant (US10946515B2) one rung up the capability ladder. The progression — from learning to grasp at all, to learning to grasp a named thing — is exactly the trajectory of useful manipulation, and fencing each rung builds a layered portfolio where competitors have to design around multiple generations of claim.
From a portfolio angle, semantic grasping sits at the intersection that matters for language-conditioned robots: the robot that takes a verbal instruction and acts on the right object. As VLA (vision-language-action) models move into robotics, a foundational claim on semantically-conditioned grasping becomes strategically central. Google fencing it early, with its core robotics-learning inventors, is a deliberate stake in that future.
Caveats. Semantic grasping is actively published, and the grant turns on the specific conditioning-and-grasp-prediction step in claim 1, not on the concept of object-aware grasping. Software-method claims also age as architectures shift. Read the independent claim for how object identity enters the grasp prediction — and read the family for the full fence.
For the file: an object-conditioned learned-grasping grant extending Google's manipulation-learning chain into the language-conditioned era. Pull US11717959B2 (and the 2021 sibling) on PatentBear, read claim 1's semantic-conditioning limitation, and track the family as VLA robotics matures.