Definition
Embodied AI is artificial intelligence that learns by acting on the physical world rather than by reading about it — a model whose training signal is what actually happened when its own body moved. It differs from a chatbot in that its mistakes cannot be retried, and from ordinary robotics in that the behaviour is learned from data rather than written down as a trajectory by an engineer.
That second distinction is the one most people arrive looking for after watching a humanoid demo. A conventional industrial arm is already extremely capable: it welds the same seam six thousand times a shift with sub-millimetre repeatability, and it does so because a person specified the path. Embodied AI is the attempt to replace the specification with a policy — a function from camera pixels and a spoken instruction to motor commands — so that the same machine can handle the object nobody enumerated. Everything hard about the field follows from three consequences of trying that with a body attached.
The data cannot be downloaded. Text and images already exist; robot experience does not, and it accrues at exactly one second per second. The action cannot be undone. A wrong token costs a retry; a dropped box puts the robot in a state it was never trained on and cannot reset itself out of. The answer is due on a physical deadline. A language model that takes an extra 200 ms is slow; a balance controller that takes an extra 200 ms has already fallen over. The rest of this page is those three constraints and the arithmetic they impose.
How It Works
The data problem, in hours
The clearest way to see the scale asymmetry is to look at what a serious collection effort actually yields. DROID, presented at Robotics: Science and Systems in 2024, is one of the largest open manipulation datasets ever gathered in the wild: 76,000 human-teleoperated trajectories, 350 hours of interaction, 564 distinct scenes and 86 tasks, collected over twelve months by 50 data collectors operating 18 robots across 13 institutions in three continents.
Do the division. 350 hours over 76,000 trajectories is 16.6 seconds per episode — robot demonstrations are short. And 350 hours spread over twelve months is under an hour of usable robot experience per calendar day, from fifty people. A single robot left running continuously for a year would accumulate 8,760 hours; you would need twenty-five DROIDs to match one such robot-year.
Now compare the other side. Llama 3 was pretrained on over 15 trillion tokens (Meta, April 2024), and nobody had to physically move anything to obtain them. There is no aggregation strategy that closes a gap of that shape, because the constraint is not storage or willingness to share — it is wall-clock time on hardware. The field's largest act of aggregation, Open X-Embodiment (October 2023), pooled 60 existing datasets from 34 laboratories into over a million episodes spanning 527 skills and 22 different robot embodiments, from single arms to bimanual setups to quadrupeds. That is an enormous coordination achievement and it is still a corpus you could collect in a few thousand robot-hours. This is why "robot foundation models" are built on top of vision-language models trained on the web: the web knowledge is borrowed, because it cannot be earned.
Borrowing knowledge: vision-language-action models
A vision-language-action (VLA) model is a vision-language model whose output head has been retargeted from text to robot commands. The bet is transfer. RT-2 (Google DeepMind, 2023) tested it directly and reported 62% success on tasks involving previously unseen objects against 32% for RT-1, which was trained on robot data alone — roughly a doubling, on the exact axis the robot data was too small to cover. The backbone already knows what a stapler is, that "the extinct animal" means the plastic dinosaur, and that spilled things get wiped. None of that is learnable from 350 hours of arm motion.
The latency budget, which physics sets
Here is where having a body stops being a metaphor. Dexterous manipulation policies target 50 Hz, which means a command is due every 20 ms. That number is not negotiable the way a chatbot's p99 is; it is what the controller needs to stay stable through contact.
Against that budget, measure the model. In the real-time chunking work from Physical Intelligence (Black, Galliker and Levine, June 2025), the 3-billion-parameter π₀ VLA spends 46 ms on KV-cache prefill alone, before a single denoising step runs — already 2.3 control periods late with no answer produced. Running π₀.₅ with five denoising steps gives a 76 ms end-to-end model latency: 3.8 control periods. Google DeepMind's Gemini Robotics 1.5 report (October 2025) puts end-to-end observation-to-action latency at roughly 250 ms, or 12.5 control periods.
No amount of inference optimisation turns 76 ms into 20 ms while keeping a 3B-parameter backbone, so the field does not try. It changes what the model outputs. π₀ emits an action chunk of 50 future commands — at 20 ms each, exactly one second of motion per forward pass. The model now runs at about 1 Hz while the robot runs at 50 Hz, and Gemini Robotics reports an "effective control frequency" of 50 Hz for the same reason, splitting the system into a cloud-hosted backbone and a small decoder running on the robot itself.
The bill for that trick is open-loop execution: for up to a second, the robot is carrying out a plan computed from a picture of the world that is now a second old. That is why the interesting engineering is in the seam. Real-time chunking generates the next chunk while the current one is still playing, freezing the actions already committed to and "inpainting" the remainder so the two join smoothly. Tested with +100 ms and +200 ms of latency injected deliberately, it kept working on tasks as unforgiving as striking a match, while the naive alternative oscillated hard enough to trigger the robot's own protective stop.
Manufacturing experience in simulation
Since real experience is rationed by the clock, the field manufactures it. Physics simulators run thousands of copies of a robot in parallel on one GPU, and the speedups are not incremental: Rudin, Hoeller, Reist and Hutter (Learning to Walk in Minutes, CoRL 2021) trained the ANYmal quadruped to walk on flat ground in under four minutes and on uneven terrain in twenty minutes, on a single workstation GPU, then transferred the policies to the real machine. Twenty minutes against DROID's twelve months is the entire argument for simulation in one comparison — and it explains why reinforcement learning, which is impractically sample-hungry on real hardware, is the default method for locomotion.
The catch is that the simulator is wrong. Friction, contact, deformation, sensor noise and actuator delay all differ from reality, and a policy that finds a quirk of the physics engine will exploit it. The standard mitigation is domain randomisation — training across randomised masses, friction coefficients, lighting and latencies so the policy must work across a band that brackets the real world instead of fitting one idealised version of it.
Real-World Applications
Warehouse stowing, where touch was the unlock. Amazon's Vulcan, announced in May 2025 and running in fulfilment centres in Spokane, Washington and Hamburg, is built around force-feedback sensing rather than better cameras: it measures how hard it is pushing and how firmly it is holding, and adjusts in real time. Amazon states it can pick and stow approximately 75% of the item types it stocks. The reason this is an embodied-AI story rather than a computer vision one is that whether a grasp holds is decided by forces no camera can observe. Perception told the robot what the thing was years ago; contact is what was missing.
Commercially useful generalist policies, in unmodified settings. Physical Intelligence's π*₀.₆ (17 November 2025) was run on three real jobs rather than benchmarks: making espresso drinks in a café from 5:30am to 11:30pm, folding 50 novel laundry items in a home the robot had never seen, and assembling and labelling 59 boxes for chocolate packaging in a factory — reported at over 90% success on all three after training. The interesting part is the training recipe, not the demo: adding expert corrections and reinforcement learning on the robot's own autonomous trials to plain imitation learning more than doubled throughput and cut failures by 2× or more on the hardest tasks.
Legged locomotion that ships. Quadrupeds doing inspection rounds in substations and refineries run controllers of exactly the kind trained in the twenty-minute simulation runs above. Locomotion transferred from simulation earlier and better than manipulation did, for a concrete reason: the contact dynamics of a foot on ground are repetitive and comparatively easy to model, while the contact dynamics of fingers on a deformable object are not.
Challenges
Evaluation costs robot-hours, and this quietly shapes the whole field. You can score a language model on ten thousand held-out prompts in an hour, for cents. Scoring a manipulation policy means running it at 1× real time with a human resetting the scene between attempts. The statistics are unkind: distinguishing a 40%-success policy from a 50%-success one at the usual thresholds needs on the order of 385 trials per policy — about 770 runs, or roughly 13 hours of robot time at one minute per trial including reset, to resolve a ten-point difference. Every ablation, every hyperparameter, every seed carries that price, which is why published real-world results often rest on a few dozen trials and why so much of the field's progress is argued over in simulation instead.
Simulation rankings do not transfer, and now there is a number for it. The RoboDojo benchmark (July 2026) evaluated 30 generalist policies across 42 simulated and 18 real-world tasks. The best real-world performer, π₀.₅, reached 12.8% overall success where expert human teleoperation reached 100%. Worse for anyone shopping by leaderboard: one policy placed second in simulation and scored 0% on the real tasks, while another placed fifteenth in simulation and reached 7.2% real success. Simulated performance is evidence about a policy; it is not a forecast.
Errors compound, and imitation learning makes it worse. A task made of twenty dependent steps at 95% per-step success finishes 0.95²⁰ ≈ 36% of the time; at 99% per step the same task finishes 82% of the time. The four-point difference per step is a forty-six-point difference in outcome, and this is why "tidy the kitchen" remains hard while "pick up the mug" does not. The mechanism behind the compounding is specific to embodiment: a policy trained purely on expert demonstrations has, by construction, never observed the states its own mistakes produce, so the first small error moves it off the data distribution and the second is worse. Training on the robot's own failures — the RECAP recipe behind π*₀.₆ — attacks precisely this.
Data collected on one robot does not automatically help another. Cross-embodiment transfer is the premise of Open X-Embodiment's 22 embodiments, and it is a research bet rather than a settled result. Different arms have different kinematics, different gripper geometries, different control rates and different camera placements, so an "episode" from one is not obviously the same kind of object as an episode from another.
The hardware has a maintenance term that software does not. Actuators wear, gearboxes backlash, force sensors drift, and a policy calibrated against a fresh machine degrades as the machine ages. There is no equivalent failure mode in a model served from a data centre, and no software update fixes a worn harmonic drive.
Future Trends
Reinforcement learning on the robot's own experience, rather than more demonstrations. The trajectory of the last two years runs from imitating teleoperators, to imitating with corrections, to learning from autonomous trials. This is the field's only route to data that does not cost a human operator per hour, and π*₀.₆'s reported doubling of throughput is the first strong commercial-setting evidence that the route works.
The two-rate architecture hardening into a standard. A large backbone that reasons in the hundreds of milliseconds, an action chunk covering roughly a second, and a small decoder on the robot closing the fast loop — this shape now recurs across independent labs, and the open research question has narrowed to how much staleness a chunk can absorb before the world has moved out from under it. Whether the backbone lives on the machine at all is a separate edge AI decision, and Gemini Robotics answers it both ways at once by splitting the model across the network boundary.
Evaluation as shared infrastructure. Because nobody can afford a private evaluation fleet at 13 hours per comparison, benchmarks are starting to ship the robots with the tasks — RoboDojo includes a real-world evaluation system with remote cloud access. If that takes hold, "we tested it on the real thing" stops being a claim only the best-funded labs can make.
Human video as a data source, with an unresolved problem underneath. There are far more hours of people manipulating objects on the internet than there will ever be hours of teleoperation, and learning from them sidesteps the collection bottleneck entirely. The obstacle is that the video records what moved, not what forces were applied, and — per Vulcan — force is frequently the thing that decided the outcome.