The failure mode this application quietly fences against is the one every home and service robot eventually hits: an instruction that only makes sense if the robot remembers what it already did. "Put it back" is meaningless without a record of where "it" was; "hand me the cup I asked for" needs to know which cup, and who asked. On June 25, 2026 a patent application published that is directed exactly at that gap. Titled "Method And Device With Robot Control," it is assigned to Samsung Electronics Co., Ltd. and carried at US20260178047A1. Before anything else, the label that governs how to read it: this is a published application, not a granted patent. It records what Samsung filed and asked an examiner to consider, not anything it can yet enforce. With that fixed, the question is what claim 1 actually covers.

Read the independent claim, not the title. Claim 1 is a processor-implemented method with two moves. First, it updates a dynamic scene graph by adding, to that graph, object-level history information for a target object "of which a state changes according to an interaction with a robot that occurs through an execution of an instruction for a task of a target robot." Second, it controls the target robot based on the updated scene graph, in response to receiving a target instruction corresponding to the target object. The structural novelty is not the scene graph itself — graphs of objects and their spatial relations are well-worn in robot perception — but the binding of per-object history into the graph and the use of that history as the control input for a later instruction. The graph stops being a snapshot of the present and becomes a record of what the robot changed.

A processor-implemented method comprising: updating a dynamic scene graph by adding, to the dynamic scene graph, object-level history information corresponding to a target object, of which a state changes according to an interaction with a robot that occurs through an execution of an instruction for a task of a target robot; and controlling the target robot based on the updated dynamic scene graph, in response to receiving a target instruction corresponding to the target object.— Method And Device With Robot Control, US20260178047A1

What the dependent claims actually add

The dependent claims are where the scope gets specific, and they read as a deliberate enumeration of what "history" means and how it is stored. Claim 2 defines the history information as any combination of a previous state of the object — previous position coordinate, previous direction, positional relationship with other objects in the same space — plus a first instruction previously performed on the object, time information for that instruction, and the previous user who issued it. So the record is not just "where the cup was" but "where it was, when it last moved, and who asked." Claim 3 extends the history beyond the target object to co-changed objects: when moving one object changes the state of others, history of those others is added to the object node as an attribute. Claim 5 is the data-structure limitation — the history is stored "in a plurality of slots of an object node" as a stack, which fixes a last-in-first-out ordering of an object's recent states. These are the limitations that turn an abstract "remember things" into a claimed mechanism an examiner can compare against prior art.

Two further dependents broaden the setting in ways worth flagging because they are where the claim reaches past a single robot. Claim 6 is directed to the multi-robot case: when another robot is present in the space where the target robot executes the instruction, the graph is updated with information on which robot performed the action, and the updated graph is shared with the other robot. That is a fleet-coordination limitation hiding inside a single-robot disclosure — a shared world-memory across robots, attributed by actor. Claims 7 and 8 add a voice path: the system recognizes the voice of the user issuing the instruction, writes that user into the object's history, and then controls the robot to execute the instruction for the recognized user. Claim 9 closes the loop on why the history exists at all — it supplements "a missing information item that is missing from the target instruction" from the updated graph, then controls the robot on the completed instruction. That is the claim that turns "put it back" into a resolvable command: the missing referent is filled from object history.

Where this lands in the patent landscape

The classification is the orienting fact. The application's main CPC is G05D 1/648 — within G05D 1, the family for control of position, course, or altitude of land, water, air, or space vehicles and, increasingly, autonomous mobile robots. The 1/648 placement points at control of movement using stored environmental or map-like representations rather than at the manipulation-mechanics class (B25J) where end-effector and joint inventions live. That is consistent with the disclosure: the claimed contribution is a control-and-memory architecture, not a gripper or an actuator. In landscape terms, the application sits at the intersection of scene-graph perception and instruction-grounded control — the layer where a natural-language or voice command has to be resolved against a remembered world state before any motion is planned.

Read against Samsung's own recent record, this hero application is not a standalone; it published into a visible cluster of robot-control filings, each directed at a different slice of the same problem. US20260178044A1, "Robot Control Apparatus And Method Thereof," is directed to path selection: determining a target region on a map that includes a region of interest, then choosing a driving path by comparing an "entry score" at a given time point against a reference score. That is the movement-planning companion to the scene-graph memory of the hero filing — one decides where to go, the other decides what the robot knows when it gets there. US20260175429A1, "Robot Navigation Method, Electronic Device, And Computer-Readable Storage Medium," is directed to navigating from an initial map annotated with obstacle-area types, using real-time ultra-wideband localization to identify which obstacle area the robot occupies and applying an area-type-specific navigation parameter — a safety-oriented localization-and-navigation layer beneath the instruction layer.

The cluster reaches into manipulation mechanics as well. US20260175431A1, "Robot With Force Control," is directed to an arm assembly using low-ratio drives on its primary axes and to determining external Cartesian forces on the end effector from a motor's torque change — explicitly "without use of force sensor information." That is a sensorless-force-estimation claim, a different CPC neighborhood from the G05D hero but part of the same recent robot-control filing wave. Taken together, the four records sketch a stack: navigation and localization (US20260175429A1), path selection (US20260178044A1), force-aware manipulation (US20260175431A1), and — at the top — an instruction-grounding memory that lets a later command be resolved from what the robot already did (US20260178047A1). Whether the claims that ultimately issue track the published language is a question for prosecution, and the scope an examiner allows may be narrower than the dependents read today. For the purpose of reading the record as filed, the coverage is plain on its face: a dynamic scene graph carrying per-object, per-user, time-stamped history, used as the control input when a new instruction about that object arrives.