The boring robot already shipped; the next problem is making several of them work together. US12263599B2, granted to Intel in April 2025, covers "Collaborative multi-robot tasks using action primitives."

Classified under B25J 9/1669, B25J 9/163 (manipulator control) and G05B 13/027 (adaptive control), the patent's idea is a shared vocabulary. Instead of programming each cooperative task from scratch, the system composes it from 'action primitives' — reusable basic moves like grasp, place, hand-off, hold — that multiple robots can sequence and coordinate around. Complex teamwork becomes a composition of simple, shared verbs.

“Various aspects of methods, systems, and use cases include techniques for training or using a model to control a robot.”— U.S. Patent No. 12,263,599 source

The claims sharpen "shared vocabulary" into something testable. Claim 1 requires an action primitive library in which each primitive defines "a control action for a robot or a cooperative action of two or more robots" — and crucially, the claimed set must include at least one of each: a single-robot control primitive and a multi-robot cooperative primitive. That distinction is the patent's center of gravity. A library of solo moves is just a macro list; what is being fenced is a library where some primitives inherently bind two or more robots together, so that "hand-off" is a first-class action rather than an emergent accident of two independent controllers. The system receives information describing a collaborative task, decides which primitive (specifically a cooperative one) to pull from the library, and trains a model to control at least one robot to complete the task using that primitive.

How the model gets trained is spelled out. Claim 4 names the methods: "a deep reinforcement learning technique or a kinematics-control technique." Claim 9 describes the RL machinery concretely — collecting next states and rewards into a replay buffer after each training step, then computing "gradients of a loss function for updating the reward." Claims 2–3 define the state and action spaces the way a reinforcement-learning practitioner would: the state space carries perception inputs, the robot's own position and orientation, the positions and orientations of the other robots, and the pose of an object in the environment; the action space is "an available action primitive or a goal location vector." So a robot does not output raw joint torques during planning — claim 5 has each robot emit a tuple of "an object, a primitive, and a goal vector." The primitive is the unit of decision.

The reward structure is where coordination is actually encoded, and the claims are unusually explicit about it. Claim 6 allows "a global reward for all robots"; claim 7 allows "a plurality of rewards for each subset of robots or each individual robot"; and claim 8 specifically allows "a negative reward corresponding to collision avoidance." Reading those together, the patent is not hand-waving about teamwork — it is claiming the standard multi-agent RL knobs (shared vs. individual credit, plus an explicit penalty for robots hitting each other) wired to a primitive-based action space. Claim 12 adds a hard constraint that movement of each robot "is restricted to only movements corresponding to actions or interactions in the set of action primitives," which is what makes the whole scheme tractable: you are not searching continuous joint space for two arms at once, you are sequencing from a finite, vetted vocabulary.

ROI per square foot, not per keynote, and the multi-robot ROI is in coordination overhead. Two robots that cannot cooperate are just two single-robot cells; two that can hand off work, share a workspace and synchronize unlock tasks neither could do alone. A primitive-based vocabulary is what makes that coordination programmable rather than bespoke per task. The library is also designed to grow: claim 14 lets the system identify a sub-task made of two or more primitives and store it back in the database as a new entry — composition that compounds, so yesterday's two-step hand-off becomes today's single callable verb. Claim 13 notes the primitives themselves can be authored either by reinforcement learning or by classical kinematics control, so the library mixes learned and engineered moves.

The honest difficulty is the coordination itself — timing, collision avoidance between the robots, and what happens when one primitive fails mid-sequence while another robot is depending on it. A shared action library makes composition easier but does not erase the hard real-time choreography of bodies sharing a space. The collision-avoidance reward in claim 8 is a training-time signal, not a runtime guarantee; it shapes a policy toward not colliding, which is not the same as proving it won't. The patent is about the vocabulary; the conducting is still hard.

The deployment architecture in claim 10 is a strategic tell. The processing circuitry and memory can sit in "a command device in communication with each of the set of robots," and the trained model is then deployed "to control actions of the set of robots… independent of the command device." Train centrally; run decentralized. That is the shape of a platform: a coordination layer that learns the choreography and then pushes it down to the robots so they can execute without a conductor in the loop. Like Nvidia in autonomy, Intel is positioning as infrastructure — the coordination and compute layer beneath robot fleets rather than a robot maker itself. Action primitives are a platform play: define the vocabulary, and you sit underneath everyone who speaks it.

For readers tracking warehouse and factory automation, multi-robot coordination is the next frontier after single-robot competence. The deployments that scale are fleets that cooperate, and the IP behind them — shared primitives, composable tasks, per-robot and global reward shaping, a train-centrally/run-locally split — is where the quiet, durable value in industrial robotics is accumulating.