The boring robot already shipped; the expensive part is teaching it the next job. US10773382B2, granted to Google's X Development in September 2020, goes straight at that cost. "Machine learning methods and apparatus for robotic manipulation and that utilize multi-task domain adaptation" is about not starting over.

Here is the unit-economics problem the patent solves. A learned grasping model trained on Task A is useless for Task B if you have to collect a fresh mountain of data for B. Domain adaptation — classified under B25J 9/163 with the G05B 2219 learning-control annotations — lets the model carry most of what it learned from A into B, including the gap between simulation and the real robot.

“Implementations are directed to training a machine learning model that, once trained, is used in performance of robotic grasping and/or other manipulation task(s) by a robot.”— U.S. Patent No. 10,773,382 source

The training scheme in claim 1 is more specific than "transfer learning," and the structure is the invention. It draws on two streams of examples. Simulated training examples come from a simulated robot manipulating simulated objects, with the example output keyed to "whether a first robotic manipulation task was successful." Real training examples come from a physical robot manipulating real objects, with output keyed to "whether a second robotic manipulation task… distinct from the first." Note the asymmetry the claim insists on: the simulated and real examples are deliberately for different tasks. The model processes the simulated input and computes a first loss against simulated success; it processes the real input and computes a second loss against real success; and it trains "at least the portion of the machine learning model based on both the first loss and the second loss." The cheap, plentiful simulation teaches one task; the scarce, expensive real-robot data adapts the shared portion to the real world via a different task.

The third loss is what makes the bridge work, and it is the technical core. Claim 4 introduces "a third loss based on intermediate output generated during generating the second predicted output," and claim 6 names it: "a domain-adversarial similarity loss." This is the sim-to-real glue. An adversarial similarity loss pressures the model's internal features to look the same whether the input came from simulation or reality — so a classifier cannot tell which domain a feature vector came from. If the features are domain-invariant, skills learned in simulation transfer to the real robot without the model fixating on rendering artifacts that don't exist in the real world. Claim 5 says the third loss modulates "an effect, of the second loss," and claim 7 describes the layered training: train only the shared portion on the real (second + adversarial) losses, while training the entire model on the simulation loss. Simulation teaches the whole network; reality fine-tunes the shared backbone and forces its features to generalize across the domain gap.

The claims even name the two manipulation tasks being bridged. Claim 13 and claim 15 specify that the simulated/first task is "instance grasping" — grabbing a specific target object, which claim 17 supports with a segmentation mask "tailored to a target object" — while the real/second task is "indiscriminate grasping," grabbing any object, for which the real input "lacks any segmentation mask." So the concrete story is: learn the hard, target-specific skill cheaply in simulation where you can render perfect masks, learn the easy, any-object skill from real robots where masks are expensive, and use domain adaptation to fuse them. Claims 11–12 even allow the two input streams to have different dimensions handled by an "additional machine learning model" — an indiscriminate-grasping model that shares a backbone with the instance-grasping model but has fewer input dimensions. The architecture is built to mix two task types with two data sources.

ROI per square foot, not per keynote, and the ROI math is brutal without transfer. If every new SKU, every new bin geometry, every new gripper requires its own training campaign, automation never pencils out for a high-mix warehouse. Domain adaptation is the lever that turns a one-task demo into a many-task deployment. And the payoff is claimed: claim 2 closes the loop by "using the machine learning model in control of an additional real robot in performance of the first robotic manipulation task," with claim 3 describing the runtime as "iteratively applying inputs… and servoing an end effector… based on the iterative predictions" — the closed-loop visual servoing that this lineage of grasping work is known for.

The honest limit the patent implies is that adaptation is not free transfer. The new task has to be close enough to the old one for the learned features to apply; a robot that learned to pick boxes does not automatically learn to fold laundry. The domain-adversarial loss aligns features across the sim-to-real gap, not across arbitrarily different tasks — instance grasping and indiscriminate grasping are siblings, which is why they share a backbone. The technique compresses the cost of related tasks, not unrelated ones.

This grant is part of the X lineage that became Intrinsic, Alphabet's industrial-robotics arm, and the multi-task thinking here shows up directly in Intrinsic's later grasping patents; the companion grant US10639792B2 ("Deep machine learning methods and apparatus for robotic grasping") sits in the same family. The 2020 filing is the conceptual seed: treat manipulation skill as a reusable asset, not a per-task expense.

For anyone evaluating an automation vendor, the question this patent frames is the right one to ask: not "can your robot do this task," but "what does it cost you to teach it the next one." The patent's answer — train in simulation, adapt to reality with an adversarial feature-alignment loss, and share a backbone across related grasping tasks — is a concrete recipe for driving that number down. That number, not the demo, is where the deployment lives.