Danijar Hafner’s World-Model Agents Face Their Hardest Test: The Unexpected

Danijar Hafner has moved his world-model agents from computer games into a stealth startup filled with humanoid robots, despite keeping its product largely undisclosed. The company, reportedly called Embo, represents a direct test of his central research idea. An agent should rehearse possible futures before acting in the real world.

That approach has already produced striking results inside controlled environments. Hafner’s Dreamer research trained agents through imagined experience, including long sequences required to collect diamonds in Minecraft. Physical robots create a harder problem because faulty predictions can cause real damage.

The timing also raises the stakes. Google DeepMind, Hafner’s former employer, is expanding Gemini Robotics across humanoids and other machines. Other well-funded robotics companies are pursuing general-purpose control models. Hafner is betting that prediction, planning, and adaptation will matter more than memorizing larger collections of demonstrations.

Danijar Hafner’s World-Model Agents Leave the Screen

The new venture turns a research program about imagined futures into a physical robotics company.

According to the original startup profile, Hafner’s sparse office sits in San Francisco’s SoMa district. The company remains in stealth mode, without even a name displayed on its door.

Humanoid robots reportedly occupy the unfinished space. They come in different shapes and sizes, with several imported from China. These machines provide physical bodies for the planning systems Hafner has spent years developing.

Hafner left Google DeepMind in the fall of 2025, according to the report. His departure shifted the Dreamer research line from a major laboratory into an independent commercial effort. Public reporting and industry directories identify the startup as Embo, although its website, product scope, and business model remain unclear.

The company’s stated challenge is easier to describe than to solve. A robot entering an unfamiliar home cannot rely on an exact recording of every room. It must understand new furniture, displaced objects, changing light, people, pets, and incomplete instructions.

A robot trained only to reproduce familiar movements can fail when one condition changes. It might reach toward the expected location after someone moves a cup. It might continue a walking path after a chair blocks the route.

Hafner’s alternative gives an agent an internal predictive model. A world model is a learned representation that estimates how an environment changes after an action. The agent can use those estimates to evaluate possible behavior before moving.

That process differs from generating a single long plan and following it blindly. A capable physical agent must predict, act, observe the result, and update its plan. Each new observation can reveal that its previous internal forecast was incomplete.

The startup therefore rests on a demanding proposition. World models must remain useful when reality differs from the training data. They also need to operate quickly enough for a robot to respond while events unfold.

Dreamer provides the research foundation for that proposition. Hafner and his collaborators began developing the family before the latest wave of language-model agents. Their work focused on learning behavior through interaction, rather than predicting text alone.

The first Dreamer system learned long-horizon behaviors from visual input by practicing inside a compact internal representation. DreamerV2 extended the approach across Atari games. DreamerV3 sought to make the same basic algorithm work across many different domains.

That progression matters because Hafner is not entering robotics with an untested conceptual slogan. He is applying a documented research program to a much less forgiving setting. The open question concerns how far its past results transfer.

A game exposes an agent to defined actions and measurable rewards. A humanoid robot encounters noisy sensors, worn components, uncertain friction, and objects with hidden properties. People also behave less predictably than simulated characters.

Embo’s early significance comes from that transition. The company is not merely building another humanoid body. It appears focused on the intelligence layer that decides what a body should do when its surroundings change.

Why Imagined Experience Matters Now

Robotics companies need agents that learn from limited physical experience without treating every mistake as a required lesson.

Modern AI models have benefited from vast digital datasets. Robotics lacks an equivalent supply of clean, diverse, action-labeled experience. Recording physical movement costs time, requires hardware, and introduces safety risks.

A language model can process another document without dropping it on someone’s foot. A robot collecting new training data can collide, damage an object, overheat a joint, or require a human reset. Those constraints make physical trial and error unusually expensive.

Simulation helps, but simulated worlds simplify reality. Textures, contact forces, sensor noise, deformable materials, and human behavior can differ from their physical equivalents. A policy that works in a simulator can fail after deployment.

World-model agents offer a possible middle path. The system learns from real observations, then generates internal predictions about what might follow. Training can use those imagined trajectories without executing every candidate action on hardware.

Hafner’s earlier DayDreamer project tested that principle on physical machines. Its robot learning paper described agents learning directly in the real world without relying on simulators.

The researchers evaluated DayDreamer across four robots. Tasks included locomotion, object handling, and visual navigation. A quadruped learned to recover, stand, and walk from experience gathered by the physical machine.

Those experiments did not establish a general-purpose household robot. They did show that learned world models could reduce dependence on extensive simulation and handcrafted task engineering. That evidence now supports Embo’s larger wager.

DreamerV3 expanded the case for a reusable algorithm. The system used one configuration across more than 150 tasks, according to the peer-reviewed DreamerV3 study. Its domains included continuous control, Atari games, and Minecraft.

The Minecraft result became an especially visible demonstration. Collecting a diamond requires exploration, resource gathering, tool construction, and many dependent actions. Rewards remain sparse because most intermediate movements produce no immediate success signal.

DreamerV3 learned the task from pixels without human gameplay data or a staged curriculum. The result suggested that imagined trajectories could support distant goals, not only immediate reactions.

That ability connects directly to physical agents. Cleaning a room involves many steps whose value appears only later. A robot might first move an obstacle, open a container, retrieve an item, and then place it correctly.

A reactive system can complete each familiar motion yet still lose the broader objective. Long-horizon planning requires tracking how early choices change later possibilities. Hafner’s work treats predictive models as the mechanism for preserving that connection.

Dreamer 4 pushed the idea further. It trains behavior within a scalable world model that simulates complex interactions. The system reportedly completed Minecraft’s diamond task from offline experience, without interacting with the game during training.

The Dreamer 4 project says the task involves sequences exceeding 20,000 mouse and keyboard actions. It also reports using 100 times less data than OpenAI’s earlier Video PreTraining agent.

Those are research results within a game, not evidence that a humanoid can safely handle an unfamiliar kitchen. Still, they address two pressures facing robotics developers. They target long action sequences and reduce dependence on online experimentation.

The urgency comes from a growing mismatch between robot hardware and robot adaptability. Humanoid bodies have improved, but many demonstrations still occur in prepared environments. A system can look capable while every relevant object remains carefully positioned.

Commercial deployment demands something less theatrical. A useful robot must recover after missed grasps, interpret changed surroundings, and choose safer alternatives. Those requirements make prediction under uncertainty central to the market.

Prediction Challenges the Imitation-First Robot

The main contest is between agents that model consequences and systems that primarily reproduce patterns found in demonstrations.

Much of recent robot learning builds on vision-language-action models. A VLA model connects images and instructions to physical actions. Training commonly draws on demonstrations showing how humans or other systems completed tasks.

This approach has clear advantages. Demonstrations provide useful behavior without forcing a robot to discover everything independently. Large pretrained models can also contribute visual recognition, language understanding, and broad conceptual knowledge.

However, demonstrations describe what happened in recorded situations. They do not automatically explain how the environment works. A robot can associate an image with an action while learning little about the consequences of a different choice.

World-model agents place those consequences at the center. The predictive model estimates future states under candidate actions. An actor proposes behavior, while a critic estimates which imagined outcomes best support the goal.

DreamerV3 uses this actor, critic, and world-model structure. The world model predicts outcomes, the critic evaluates them, and the actor selects actions. The components learn together from collected experience.

This mechanism gives the agent a way to practice beyond the exact trajectory it observed. It can generate alternative internal rollouts from known states. Useful imagined experience then shapes its behavior without requiring another physical attempt.

The distinction should not be exaggerated into a clean technological divide. VLA systems can include planning, feedback, and predictive components. World-model systems can also learn from demonstrations, language, or video.

The real disagreement concerns emphasis. An imitation-first strategy starts with large collections of successful behavior. Hafner’s route starts with learning dynamics that let an agent evaluate what happens next.

Google DeepMind’s competing work demonstrates how quickly these approaches are converging. Its Gemini Robotics 2 system combines language, visual understanding, planning, and motor control.

Google says the model can control humanoids from feet to fingertips, manipulate objects, and coordinate multiple robots. It also claims adaptation to new robot bodies within hours. Those statements come from Google and require broader independent evaluation.

Gemini Robotics benefits from foundation-model knowledge and Google’s research infrastructure. Hafner’s startup brings concentrated expertise in reinforcement learning through predictive models. Both efforts target machines that adjust when conditions stop matching a script.

Skild AI presents another variation. The company describes its goal as a single general-purpose robot brain that works across different bodies and tasks. Its strategy relies heavily on broad data collection and large-scale pretraining.

Physical Intelligence has pursued general robot policies trained across tasks and embodiments. Its systems connect visual and language inputs to continuous actions. The company’s demonstrations emphasize transfer across different physical platforms.

These competitors pressure Embo in two directions. They can collect more data and deploy more hardware. They can also combine planning with pretrained models, narrowing the conceptual advantage of a pure world-model identity.

Embo must therefore show more than an elegant architecture. It needs evidence that imagined training improves adaptation, safety, or data efficiency on real robots. Those benefits must remain visible against strong baselines.

A decisive evaluation would introduce changes after training. Researchers could move objects, alter room layouts, replace tools, or interrupt a task. The test would measure whether the robot replans without extensive retraining.

Success should also include recovery. A robot that completes prepared trials but freezes after one missed grasp has not mastered the unexpected. Planning becomes valuable when the first forecast proves wrong.

That requirement makes Hafner’s strategy more ambitious than teaching a robot to anticipate everything. No model can enumerate every physical event. The practical goal is recognizing prediction errors and updating behavior before they compound.

The World Model Can Be Wrong

An agent that plans inside an inaccurate model can become confidently wrong before its body makes the mistake real.

World models compress experience. Compression makes prediction efficient, but it also discards information. A detail that appeared irrelevant during training can become decisive in an unfamiliar setting.

Consider a robot carrying a glass toward a table. Its model might predict that the surface will support the object. A transparent sheet, wet patch, loose covering, or unusual edge can invalidate that forecast.

Long imagined rollouts amplify small errors. A slightly incorrect first prediction changes the state used for the next prediction. After many steps, the internal trajectory can drift far from the physical world.

This problem is often called model bias. The policy learns to exploit patterns inside the model that do not exist outside it. A virtual strategy can earn high predicted rewards while failing during real execution.

Dreamer’s actor does not search through a perfectly accurate simulator. It trains against a learned model built from prior observations. The quality and diversity of those observations therefore constrain what the agent can imagine reliably.

Unexpected situations expose that boundary. If the agent has never observed a certain interaction, its forecast might lack the relevant physics. It can estimate uncertainty, but uncertainty estimates can also be poorly calibrated.

A physical system needs several defenses. It must compare predictions with new sensor data, shorten plans when confidence falls, and hand control to safer mechanisms. It also needs constraints that prevent high-risk actions during exploration.

Hafner’s previous results establish performance on research benchmarks. They do not establish household safety, industrial reliability, or dependable interaction with people. Embo has not publicly released the evidence needed for those conclusions.

Benchmark diversity also differs from environmental openness. DreamerV3 used one configuration across many tasks, which reduced domain-specific tuning. Each benchmark still supplied rules, action spaces, and evaluation procedures.

Homes and workplaces do not provide similarly clean interfaces. A door can stick, a person can interrupt, or an object can break. Sensors can become blocked while the robot remains responsible for its movement.

Humanoid form adds further complications. Two-legged bodies have large action spaces and can fall. Hands contain many joints, while contact dynamics change rapidly during manipulation.

The cost of an incorrect prediction also varies. Dropping a soft toy differs from dropping a knife. Planning systems need risk-sensitive objectives that distinguish those cases before deployment.

Google DeepMind acknowledges this broader challenge in its robotics materials. The company describes layered safety, including collision avoidance, force limits, and low-level controllers. High-level reasoning alone cannot replace those protections.

Embo will need comparable clarity. The startup has not disclosed how its agents constrain actions, estimate uncertainty, or recover from model failures. It has also not identified its first commercial environment.

Those omissions are normal for a stealth company, but they limit meaningful evaluation. Imported humanoid platforms can accelerate experimentation. They do not reveal whether the company has solved generalization, reliability, or unit economics.

The most persuasive evidence would come from repeated tests across unfamiliar environments. Evaluators should specify failure rates, intervention rates, recovery behavior, and the amount of additional training required.

A polished demonstration offers weaker evidence. Teams can select successful attempts and prepare favorable conditions. Continuous operation exposes rare errors that short videos tend to hide.

Data efficiency also needs careful accounting. Imagined training reduces some physical interaction, but the world model still requires real experience. Compute, sensor calibration, resets, and hardware maintenance remain part of the total cost.

Dreamer 4’s offline Minecraft result strengthens the research case for learning behavior inside a predictive model. It does not settle whether offline robot data covers the physical variations needed for safe deployment.

The central risk is therefore inseparable from the central advantage. The agent can rehearse many futures without taking physical action. Yet every rehearsal inherits the blind spots of the model producing it.

Planning Ahead Requires More Than a Longer Plan

Handling surprises depends on continuous replanning, not on predicting one perfect sequence before the robot begins.

The phrase “plan ahead” can suggest a fixed checklist. Hafner’s work points toward a more dynamic loop. An agent predicts outcomes, acts briefly, observes reality, and revises its internal state.

This distinction matters because unexpected events cannot all be predicted in advance. A robot rarely needs a complete forecast of every possible interruption. It needs a useful response when an interruption invalidates its current plan.

Imagine a robot clearing a dining table. It sees a plate, cup, utensils, and container. Its initial plan assigns each object a destination and chooses an order.

A person then places a phone beside the cup. The robot must recognize that the scene changed. It should protect the phone, reconsider its grasp path, and preserve the broader cleanup goal.

A purely reactive controller can avoid an immediate collision. A language model can describe what probably happened. A capable world-model agent must connect the change to future physical consequences.

It might predict that lifting the cup first risks knocking the phone. It could move the phone, approach from another direction, or ask for help. Each choice alters the remaining task.

Temporal abstraction becomes important here. This technique divides a long task into high-level subgoals and lower-level actions. The agent can revise one subgoal without rebuilding every motor command.

Hafner has studied this problem alongside world models. His public research overview describes temporal abstraction as breaking long tasks into subgoals that support abstract planning.

For a warehouse robot, the high-level goal might involve moving a package to a loading area. Lower-level policies handle walking, balancing, grasping, and obstacle avoidance. A changed route should not erase the entire objective.

Language can contribute by expressing goals and constraints. It can tell the robot that a red container holds fragile material. Predictive models then estimate how candidate actions affect that physical object.

This combination explains why world models and VLAs will probably merge rather than remain isolated categories. Language provides general knowledge and instruction following. World models provide action-conditioned forecasts tied to the observed environment.

Memory presents another practical requirement. An agent must retain relevant observations across a long task without treating every detail as equally important. It also needs to retrieve prior failures when similar conditions appear.

That challenge resembles the context problem faced by digital agents. Developers use structured histories and AI knowledge bases to preserve evidence beyond one model response. A physical agent needs an even tighter link between memory and current perception.

However, stored experience cannot become a substitute for verification. A robot should not assume that yesterday’s room arrangement still applies. Current sensor data must correct stale memory.

The same rule applies to its imagined futures. Prediction is a hypothesis about what follows, not a fact. Effective agents should keep testing that hypothesis against the world.

This creates a useful standard for judging Embo. The company does not need to show a robot that never encounters surprise. It needs to show one that detects surprise early and recovers without dangerous improvisation.

Recovery behavior can include stopping. Refusing an uncertain action is sometimes more intelligent than completing the task. Commercial systems need calibrated thresholds for seeking human assistance.

The hardest cases will combine long-term goals with immediate safety. A robot might know the next useful action but face uncertain contact or poor visibility. It must delay progress without abandoning the objective.

That balance will determine whether planning becomes operational capability or merely impressive prediction. A robot succeeds when its internal rehearsal improves real decisions under changing conditions.

Three Signals Will Show Whether Embo’s Bet Works

The next evidence must connect world-model research to repeatable robot performance outside prepared demonstrations.

The first signal is a public technical evaluation on unfamiliar physical tasks. Embo should define what changed between training and testing. New layouts, objects, interruptions, and robot bodies would provide meaningful stress tests.

Results should include unsuccessful trials. Intervention rates, recovery times, and safety stops matter as much as completed tasks. If Embo publishes those measurements, its claim about unexpected events becomes testable.

Strong results would reinforce Hafner’s core argument. They would show that imagined training transfers beyond games and narrow robot experiments. Weak or selectively reported results would leave the central question unresolved.

The second signal is evidence of cross-body transfer. Embo’s office reportedly contains several humanoid designs, suggesting that the company does not want intelligence tied to one machine.

A general controller should preserve useful skills when cameras, limbs, joint limits, or proportions change. Some adaptation will remain necessary. The important measurement is how much new data and engineering each body requires.

Fast transfer would challenge companies whose models depend on one hardware stack or tightly curated demonstrations. Heavy retraining would weaken the claim that Embo has learned general physical structure.

Google DeepMind already frames cross-embodiment control as a major goal. Skild AI makes a similar general-purpose claim. Embo will need direct comparisons that separate architectural benefits from marketing language.

The third signal is a credible deployment boundary. Embo must eventually name the environment where its system can operate repeatedly. A home, warehouse, laboratory, or factory presents different objects, risks, and economic requirements.

A focused deployment would not diminish the broader research vision. It would create the operational data needed to improve it. Real customers also expose failure patterns that internal benchmarks miss.

The choice of market will reveal Embo’s safety strategy. Industrial sites can constrain access and standardize procedures. Homes provide more variety but create harder social and physical risks.

Readers should also watch whether the startup releases technical papers, model documentation, or reproducible benchmarks. Hafner’s previous work has been unusually open, including code for major Dreamer systems. A commercial company can change that pattern.

Transparency would help independent researchers test the world-model approach. Closed demonstrations would force the market to judge performance through company-selected evidence. Neither outcome proves success, but they create very different confidence levels.

Embo’s broader importance does not depend on winning the humanoid market immediately. The company puts a specific technical thesis under pressure. Agents need internal models of consequences if they are expected to act beyond familiar scripts.

That thesis also reaches beyond robotics. Digital agents encounter changed interfaces, missing files, conflicting instructions, and incomplete information. They likewise benefit from predicting outcomes and revising plans after unexpected feedback.

Physical machines make the standard stricter because mistakes cross into the real world. A planning error can damage hardware or endanger a person. Reliability must therefore grow alongside autonomy.

Hafner’s record gives Embo a serious foundation. Dreamer demonstrated learning through imagined experience across games and control tasks. DayDreamer brought the method onto physical robots. Dreamer 4 extended offline imagination training across very long action sequences.

The startup now faces the decisive translation. Research benchmarks define success clearly and repeat experiments cheaply. Humanoid robots encounter environments that resist both simplification and complete enumeration.

That is why Embo is more than another stealth robotics company. It tests whether an agent can build a useful internal account of the world, notice when that account fails, and adjust before harm follows.

The next few months should bring evidence on unfamiliar tasks, cross-body transfer, and an initial deployment setting. Those signals will show whether Danijar Hafner’s world-model agents can turn imagined futures into dependable physical behavior.

Until then, the most honest conclusion remains conditional. Embo has a coherent mechanism and an unusually relevant research history. It has not yet shown that its humanoids can manage the unexpected at commercial scale.

The question for developers and AI buyers is therefore concrete: will Embo publish repeatable failure and recovery data, or only carefully prepared successes? That evidence will determine whether planning through world models becomes a dependable agent architecture or remains a compelling laboratory result.