The hard word in this title is "domain-invariant," and it names the central problem of learned manipulation. Google's grant US12112494B2 ("Robotic manipulation using domain-invariant 3D representations predicted from 2.5D vision data," issued October 8, 2024; inventors include Honglak Lee, Soeren Pirk, Seyed Mohammad Khansari Zadeh) predicts a 3D representation of a scene from cheap 2.5D (RGB-D) input, and fences the representation being domain-invariant — stable across the gap between simulation and reality, and across different objects.
The mechanism turns 2.5D vision (color plus a single depth view) into a fuller 3D understanding (G06T 7/55, depth/structure from images; G06V 20/64, 3D object recognition) that a manipulator (the B25J control codes) can act on. Domain invariance is the inventive hedge: a representation that doesn't shift when you move from the training distribution to the real robot is what makes learned manipulation deployable rather than demo-only.
For the manipulation beat, this is Google fencing the perception substrate of grasping rather than the grasp itself — a level below its semantic-grasping and deep-grasping claims. If the 3D representation transfers, every downstream manipulation policy benefits. Fencing the domain-invariant representation is fencing the thing that makes sim-trained robots work in the real world, which is the whole bottleneck of scaling manipulation learning.
From a portfolio angle, this slots into Google's deep, multi-year manipulation-learning portfolio (deep grasping 2021, semantic grasping 2023, this representation work 2024 — overlapping inventors throughout). The progression fences each layer of the stack: how to grasp, what to grasp, and now how to see for grasping. That layered density is the moat — a competitor must clear multiple generations of claim.
Caveats. Domain adaptation and learned 3D-from-2.5D are intensely published; the grant turns on the specific representation-prediction-and-transfer method in claim 1, not on the concept of domain invariance. Software-method claims age with the architectures. Read the independent claim for how invariance is achieved and read the family for full scope.
For the file: a perception-substrate manipulation grant fencing the sim-to-real bottleneck, inside Google's layered manipulation portfolio. Pull US12112494B2 (with the 2021 and 2023 siblings) on PatentBear, read claim 1's invariance limitation, and treat the multi-year family as the real fence.