In June 2026, the humanoid company 1X stood up a dedicated World Model Lab and published a number that should reframe how the industry thinks about robot intelligence. Their model, 1XWM, learns manipulation from 900 hours of egocentric human video, then fine-tunes on just 70 hours of robot data to adapt to the NEO robot's appearance and kinematics.
Read that ratio again. The robot data is the thin shim. The intelligence comes from watching humans move through the physical world, and the bottleneck moves with it: away from the model, toward the data. As 1X CEO Bernt Børnich put it, you can't fine-tune your way to AGI.
The ratio is the clearest signal yet of a shift that has been building all year. Robot learning has been dominated by Vision-Language-Action (VLA) models, which map what a robot sees plus an instruction directly to an action. World models challenge that approach at the root: they learn to predict what will happen next, then plan against the prediction. At Sequoia's AI Ascent in 2026, NVIDIA's Jim Fan delivered the eulogy.
The funeral is premature, and we will say so plainly below. But the direction is real, and its consequence matters more than which architecture wins. The data world models require, calibrated, multi-sensor, temporally coherent, and verified by people who know what they are looking at, is precisely the category Avala was built to produce.
The wall that VLAs ran into
A VLA is a vision-language model with an action head bolted on. It inherits the world knowledge of internet-scale pretraining, then learns to output low-level robot commands. The lineage is impressive: Google DeepMind's RT-2, which turned robot control into language-like tokens; the open OpenVLA baseline; Physical Intelligence's π0 series, which introduced flow matching for smooth continuous control; and Figure's Helix, the first VLA to drive high-rate, whole-body humanoid control.
The problem is structural, and the research literature is consistent about it. Because VLAs learn by imitation, they compound their own errors under distribution shift. A behavioral-cloning policy copies what it saw in the demonstrations, so unfamiliar lighting, a novel object, or a slightly different starting position sends it off script, and on long-horizon tasks one early mistake cascades into total failure. Jim Fan's critique is sharper still: most of a VLA's parameters are spent on language, which makes these models, in his words, good at nouns and knowledge but poor at physics and verbs. They can name what they see. They struggle to predict what happens when they act on it.
There is a second, quieter problem. VLAs are hungry for labeled, teleoperated action data, which means a human has to drive the robot to generate every training example. That data is expensive, slow to collect, and capped by a hard physical limit: a robot can produce at most 24 hours of real interaction per day. You cannot download it from the internet.
Why prediction beats reaction
A world model does not ask, given what I see and the instruction, what action should I take? It asks, if I take this action, what will happen next? The model learns the dynamics of the physical world, predicts a future state, then searches over possible actions for the one that reaches its goal. This is classical optimal control with one modern twist: the model of the world is learned, and so is the abstract representation it plans over.
This is the bet Yann LeCun has been making since his 2022 position paper on autonomous machine intelligence: the missing piece of modern AI is a world model, a system that tells an agent what is likely, what is plausible, and what is impossible, the way common sense does for humans and animals. The failure mode of generative video shows why the objective matters. Forced to predict the exact pixels of the next frame, a model facing uncertainty (will the ball bounce left or right?) hedges by averaging, and the average of two sharp futures is a blur. The space of possible next frames at high definition exceeds the number of atoms in the observable universe. Predicting all of them is the wrong objective.
Two camps avoid it differently. Pixel-space models like NVIDIA Cosmos and Wayve's GAIA line predict future frames directly. Latent-space models like Meta's JEPA family skip the pixels and predict in an abstract representation, so the model does not waste capacity on unpredictable detail like leaves rustling on a tree. Either way, the payoff shows up in control. Meta benchmarked its action-conditioned V-JEPA 2 on a real robot arm against a leading behavioral-cloning policy, and the gap on multi-stage tasks was stark.
The pattern is exactly what compounding predicts. A reactive policy accumulates error step by step, so a small miss while grasping becomes a total failure while placing. A planner that evaluates the full trajectory before it moves does not let those errors compound. Per Meta's reported figures, the action-conditioned model reaches 65–80% success on pick-and-place with new objects in unseen environments, after fine-tuning on roughly 62 hours of robot data. It also plans in about 16 seconds what a pixel-generation approach takes 4 minutes to compute.
The whole industry is placing the bet
This is not one lab's thesis. It is a capital reallocation.
NVIDIA Cosmos ships open world foundation models pitched explicitly as the backbone for physical AI, trained on a corpus NVIDIA describes as 20 million hours of real-world interaction, environment, and driving data. Jensen Huang's framing: just as large language models revolutionized generative and agentic AI, Cosmos world foundation models are a breakthrough for physical AI.
Meta's V-JEPA 2 was pretrained on more than 1 million hours of video with a billion-parameter encoder, then taught to act with under 62 hours of robot data.
Google DeepMind's Genie 3 generates navigable, interactive 3D worlds in real time at 24 frames per second, and Dreamer 4 learned the 20,000-action Minecraft diamond task entirely from offline data, with zero live environment interaction.
Wayve's GAIA-3, a 15-billion-parameter model, generates controllable multi-camera driving scenarios for both training and evaluation.
Generalist AI's GEN-1 reports 99% success rates, trained on a dataset the company puts at half a million hours of real-world data, with roughly 99% of its parameters trained from scratch rather than fine-tuned from a language model.
The money has followed the architecture. Yann LeCun left Meta to co-found the world-model startup AMI Labs with Alexandre LeBrun, which raised a reported $1 billion in seed funding, backed by NVIDIA, Temasek, Bezos Expeditions, and Eric Schmidt. Fei-Fei Li's World Labs raised on a similar scale for spatial intelligence. The smart capital is betting that the systems which can predict the physical world will be the ones that operate in it.
Convergence, not a funeral
Now the correction. VLAs are not dead, and the eulogy was rhetorical. The most credible reading of the evidence, including NVIDIA's own technical analysis, is that the winner is neither a pure VLA nor a pure world model but a hybrid: newer policies already borrow world-model components, and many world-action models borrow the action-expert recipe that VLAs pioneered. The two lineages are converging, not replacing one another.
World models also have real, unsolved problems, and 1X is refreshingly candid about them. Its world model runs on a 14-billion-parameter video backbone, and inference currently takes about 11 seconds to imagine a single five-second rollout, so the robot has to stop and think before it acts. Some tasks, like pouring cereal or drawing a smiley face, still sit near 0% success. Monocular video gives weak 3D grounding. Pure pixel prediction is prone to what Jim Fan calls physics slop, where a model figures out that if you're not looking, geometry is optional. Latency, long-horizon fidelity, and sim-to-real transfer remain genuine engineering challenges. The paradigm is winning, but it has not finished the job.
The bottleneck moved from architecture to data
Here is the part that holds regardless of which architecture wins. Every advantage a world model has over a VLA, learning physics, planning against consequences, generalizing to new tasks, is downstream of the data it learns from. And the data world models crave is not the data the last generation needed.
World models train on a closed-loop flywheel: web-scale video, then egocentric human video, then simulation, then teleoperated robot data, then on-policy fleet data from real deployment, feeding back into the next round of training. Research systems that pair generative world models with 4D Gaussian Splatting do this explicitly to narrow the sim-to-real gap, and they find that adding real 3D structure lifts task success over flat 2D approaches. The inputs that make it work: calibrated, time-synchronized, multi-sensor streams across LiDAR, radar, camera, and IMU; 4D reconstruction of dynamic scenes; force and proprioceptive signals; and labels that are not just accurate but failure-aware and traceable.
The field's own surveys now say this directly: the bottleneck is data infrastructure, not architecture. And it is not a volume problem you can solve by scraping more. Naive pooling of mismatched robot data can cause negative transfer, where more data makes the model worse. Sam Sinha, who left Luma AI to lead world models at 1X, put the principle as plainly as anyone.
Regulators are arriving at the same requirement from a different direction. Under EU rules that take effect in January 2027, training data for high-risk AI systems must be relevant, sufficiently representative, free of errors, and complete. That is not a labeling nice-to-have. It is a legal standard for traceable, verifiable ground truth, and it favors whoever can prove where every label came from.
Why this is the moment Avala was built for
Avala is Fleet Learning Infrastructure for Physical AI, and we run the data-to-release loop end to end: collection, 4D reconstruction, auto-labeling with human verification, and versioned releases the customer trains and evaluates against. We do not treat data as something you collect and discard. We engineer ground truth, and we make it auditable, versioned, and sensor-native from the first frame to deployment.
It starts with collection. The data the world-model era needs has to be captured from the real world, not scraped from it. Avala runs real-world data collection through our own team and a partner network of homes and factories across the United States. That is the egocentric human and real-environment data these models run on, produced at a scale no single robot fleet can match, in exactly the settings, households and industrial floors, where embodied systems have to work.
From there, Avala's 4D Engine uses Gaussian Splatting to reconstruct scenes with deterministic temporal coherence across LiDAR, radar, camera, and IMU streams. That word, deterministic, is the difference that matters. World models trained on probabilistic labels inherit probabilistic errors. Ground truth that is deterministic, temporally coherent, and traceable is the substrate that lets a world model learn physics instead of learning slop.
The economics work because of how we combine machines and people. Foundation models now handle more than 80% of annotation. Our workforce guarantees the quality on the cases that matter, the edge cases and failures where a world model either learns the truth or learns a confident mistake. AI agents auto-label, human experts verify, the models retrain, and the loop repeats, moving an engagement from fully human labeling toward roughly 90% agent-automated within months.
Those annotators are coworkers, not a disposable layer. We pay above-market wages and build long-term career pathways, because high-retention teams accumulate deep context, and deep context is what produces accuracy on the hard cases. People-First AI is not charity. It is an operational necessity. The ethics and the effectiveness point in the same direction, which is the only kind of moat that lasts. I was the third engineer on Tesla's Autopilot AI team, and Avala is the data infrastructure I wished had existed then: a glass-box engine that replaces fragmented vendors with traceable ground truth.
Imitation, imagination, and the thing underneath both
The shift from imitation to imagination is the right description of where robot learning is heading: from copying what the model saw to predicting what will happen. But prediction is only as good as the reality the model was shown, and reality has to be captured before it can be learned. Compute is abundant. Hardware is scaling. Algorithms are converging. The scarce input is the physical world itself, measured: raw sensor streams turned into computable, verifiable truth. That is the work.
Whether the winning architecture turns out to be a world model, a VLA, or the hybrid the evidence points to, it will run on high-quality, traceable, real-world data. The map can finally learn to dream. What it dreams of is still drawn from the territory, and the territory has to be measured by someone who can be trusted to get it right.
The path to global abundance runs through Physical AI. The path to Physical AI runs through ground truth. The path to ground truth runs through Avala.
Emal Alwis is the founder and CEO of Avala. Building in Physical AI and want to compare notes on the world-model shift? Reach out.
