Count the methods and the story changes: DeepMind's 2025 robot-learning patent runs two at once. US12343874B2, "Reinforcement and imitation learning for a task," combines learning from demonstration with learning from reward.

The B25J filing tells a different story than the keynote. Classified under B25J 9/163 (manipulator control) with G06N 3/08 and G06N 3/045 (neural-network learning), the patent blends two approaches that are usually presented as rivals. Imitation learning copies human demonstrations; reinforcement learning optimizes a reward through trial and error. The patent uses both in one scheme.

“A neural network control system for controlling an agent to perform a task in a real-world environment, operates based on both image data and proprioceptive data describing the configuration of the agent.”— U.S. Patent No. 12,343,874 source

Claim 1 lays out the hybrid concretely, and the structure is the patent. The system first obtains, for many performances of the task "by a real-world agent controlled by an operator," a demonstration dataset for each — the imitation data, gathered by a human teleoperating the robot. It then trains a neural network to control a simulated robot using those demonstrations. At each step the network takes two input streams named in the blockquote — "simulated image data encoding simulated camera images" and "simulated proprioceptive data… characterizing configurations of the… agent" (joint angles and the like) — and emits control commands for the robot's components. For each command set it computes "a task reward value characterizing how successfully the task is carried out." Then comes the fusion: the network's parameters are adjusted "based on a hybrid energy function including (i) an imitation reward value derived using the demonstration datasets… and (ii) a task reward term computed using the task reward values." One objective, two signals — how human-like the behavior is, and how well the task is actually done.

The imitation half is implemented as an adversarial discriminator, which is the patent's most specific technical commitment. Claim 2 uses the demonstration datasets "to generate a discriminator network" and computes the imitation reward "using the discriminator network and the sets of… control commands." This is the generative-adversarial-imitation idea: a discriminator is trained to distinguish the robot's behavior from the human demonstrations, and the robot is rewarded for fooling it — for producing trajectories the discriminator mistakes for human. Claim 3 notes the discriminator can receive "data characterizing positions of one or more objects in the simulated environment," grounding the imitation signal in the scene, not just the robot's own motion. That sidesteps the brittleness of naive behavioral cloning, which only ever copies and breaks on unseen states.

The reinforcement half is the standard actor-critic machinery, claimed explicitly. Claim 5 describes computing updates "using an activation function estimator obtained by subtracting a value function from the initial task reward value" — that subtraction is an advantage estimate, reward minus a learned baseline, the core of policy-gradient methods. Claim 7 lets the value function be "computed by an adaptive model" (a learned critic). Claim 12 allows training "in parallel with… a plurality of additional instances of the neural network by respective workers" — the distributed, many-actor setup characteristic of DeepMind's large-scale RL. The network architecture in claims 8–10 is a convolutional front end on the image data, an adaptive component (claim 9: "a perceptron") fusing the convolutional output with the proprioceptive data, and optionally a recurrent layer over both — vision plus joint state, fused, with memory.

Here is why combining them is the pragmatic move. Imitation learning is sample-efficient — a few demonstrations get you started fast — but it cannot exceed the demonstrator and fails on situations the demos never showed. Reinforcement learning can discover better-than-human behavior and handle novel states, but it is data-hungry and explores dangerously at first. Each method's weakness is the other's strength. Claim 13 adds a clever bootstrapping trick that uses the demonstrations to tame RL's exploration problem: define task stages, draw demonstration states for each stage, then "randomly select[] an initial state from the… demonstration states" to start a training episode partway through the task. Instead of exploring from scratch and almost never stumbling onto success, the agent is dropped into states the human reached, so reward signal appears far sooner. Demonstrations don't just shape the reward — they seed where reinforcement gets to explore from.

The blend, stated plainly: use demonstrations to bootstrap a competent starting policy, then use reinforcement to refine and extend it beyond what was demonstrated. The robot learns the safe basics from humans and the hard edges from reward. Crucially, claim 1 trains in simulation and then closes with "using the trained neural network to control the real-world robotic agent" — train in sim with both signals, deploy on the real robot. It is the obvious-in-hindsight synthesis, and DeepMind patented a concrete way to do it.

The honest limit is reward design. Reinforcement learning is only as good as the reward function, and specifying a reward that produces the behavior you actually want — without the reward-hacking pathologies the field is famous for — is its own hard problem. The hybrid energy function softens this, because the imitation discriminator constrains the policy toward human-like behavior even where the task reward is sparse or gameable, but the task-reward term in claim 4 still rests on "computing an initial task reward value based… on the final state of the simulated environment," and defining that state-based reward well remains on the engineer. The patent combines the methods; it does not abolish the difficulty of telling a robot what "good" means.

For readers tracking embodied AI, this grant is a marker of consolidation. The early debate was imitation versus reinforcement; the mature answer, as DeepMind's filing shows, is both — a discriminator-based imitation reward and an advantage-based reinforcement term fused in one objective, with demonstrations seeding the exploration, staged so each covers the other. That is what a maturing field looks like — not a winner, but a recipe.