WAIC 2026 World Model Debate: Six Embodied AI Teams on Physics, Benchmarks, and Deployment

World Models at WAIC 2026: Where Six Embodied-AI Teams Agree—and Where They Do Not
Introduction
World models have become one of the most active areas in embodied AI, but the field still lacks a settled definition.
Researchers disagree on what a world model should represent, how accurately it needs to model physics, whether future states should be generated in pixels or latent space, how much prediction a robot should perform before acting, which data modalities matter most, and even how progress should be measured.
That lack of consensus was unusually visible at WAIC 2026.
On July 19, a forum organized around six Chinese embodied-AI companies brought together technical leaders representing different approaches to robot world models:

• Tao Dacheng, Chief Scientist of ACE Robotics, as moderator.
• Shen Yujun, Chief Scientist of Ant Group’s Robbyant team.
• Zhu Zheng, co-founder and Chief Scientist of GigaAI.
• Xia Zhongpu, co-founder and co-CTO of Wujie Dongli.
• Ren Guanghui, AI Algorithm Director at AgiBot.
• Wang Hao, co-founder and CTO of X Square Robot. The discussion was valuable precisely because the speakers did not converge on one technical answer. Some want world models to become increasingly precise representations of geometry, motion, and force. Others care more about whether a model predicts enough physical structure to choose a successful action. Some view long-horizon consistency as essential. Others argue that prediction becomes counterproductive when it consumes too much of the robot’s decision window. Some expect the field to converge around a shared representation. Others expect hybrid architectures, tactile sensing, or a standardized benchmark to become the first point of agreement. Behind these disagreements is a more practical question: What does a world model need to do before it becomes useful outside a demo? Three Broad Technical Routes The six companies do not use identical architectures, but the panel discussion can be roughly organized into three broad approaches based on where the model predicts the world. Route Representative approach Main idea Pixel-space generation GigaAI Predict future visual states directly or through generative visual representations Latent-space prediction Wujie Dongli’s JEPA-oriented direction Predict compact internal representations rather than full pixels Hybrid or unified architectures ACE Robotics, Robbyant, X Square Robot, AgiBot Combine generation, latent prediction, explicit spatial reasoning, VLA control, or simulation-oriented models This classification is useful, but it should not be treated as a permanent industry taxonomy. Several of these companies are already combining multiple mechanisms. The more important disagreement is what each group believes a useful model must preserve. What Capability Matters Most? The first major disagreement is about the world model’s relationship with physics. View 1: Physical State Prediction Should Be Accurate Xia Zhongpu argued that evaluation should include the model’s accuracy in predicting geometric state, motion, force, and related physical quantities. This view treats a world model partly as a structured predictor of physical state. For a robot deciding how to push, catch, lift, navigate, or manipulate an object, the relevant questions may include:
• Where will the object move?
• Will it collide with something?
• How will its orientation change?
• How much force will be required?
• Will the grasp remain stable?
• What state will the environment enter after the action? A latent representation can make explicit physical regression attractive because the model does not need to render every future pixel before estimating these quantities. The benefit is precision and control. The difficulty is that the physical world contains variables that are hard to observe or infer directly, such as friction, hidden mass distribution, compliance, contact state, and material properties. A model may therefore be physically structured without becoming a complete simulator. View 2: Physical Plausibility Can Matter More Than Exact Physics Shen Yujun offered a different standard. A robot does not necessarily need to reproduce every parameter of the physical system. It needs to understand the differences that matter to the action it is about to take. For example, it may need to recognize that two materials respond differently when dropped, squeezed, dragged, or grasped without estimating every coefficient in advance. This is closer to action-relevant physical understanding than full physical simulation. The distinction can be summarized as: Exact physical simulation: Predict the detailed state of the world as accurately as possible.

Action-relevant physical reasoning:
Predict enough of the world to choose a safe and effective action.
The second target can be computationally cheaper and may be sufficient for many real robot tasks.
It also matches the design of systems that grew out of video-generation models. These models naturally learn which future visual outcomes look plausible, even when they do not explicitly expose all underlying physical variables.
The Real Question: What Error Changes the Action?
The panel’s disagreement suggests a more practical evaluation rule:
A prediction error matters most when it changes what the robot should do.
If a robot can successfully place a cup without estimating the exact friction coefficient of the table, that missing number may not matter.
If the same missing physical understanding causes the cup to slip, then the error is important.
The required physical fidelity therefore depends on:

• The task.
• The safety margin.
• The robot embodiment.
• The action horizon.
• The cost of failure.
• The time available for inference. Long-Horizon Consistency or Fast Decisions? The second disagreement concerns how far into the future a world model should predict. AgiBot: Long-Horizon Physical Consistency Matters AgiBot has invested heavily in simulation and world-model-based evaluation. Its public AgiBot Digital World framework provides high-fidelity simulated environments, automatically generated expert trajectories, and model evaluation tools. AgiBot later introduced the EVAC world-model framework and EWMBench, extending the idea of action-conditioned generative simulation. For a world model used as a simulator, long-term consistency is important. A simulation becomes less valuable if:
• Objects gradually change identity.
• Geometry drifts.
• A robot forgets earlier interactions.
• Long sequences violate physical state.
• Multi-view outputs become inconsistent. From this perspective, the world model should preserve coherent state across a meaningful horizon. X Square Robot: Long Rollouts Can Become a False Goal Wang Hao challenged the assumption that longer prediction is always better. X Square Robot uses an end-to-end embodied-model direction in which perception, reasoning, world modeling, and action generation are tightly connected. Its public materials describe the WALL series as moving toward a unified physical-world model rather than treating world prediction and robot control as isolated systems. In such a system, the world model is not only a simulator. Prediction is part of the control loop. A longer rollout can therefore introduce a cost: More predicted future → more inference → more latency → less time left to act If the robot’s environment changes faster than the model can complete its reasoning, a beautifully consistent long-term prediction may be operationally useless. For control, a shorter prediction that arrives at the right time can be more valuable than a longer one that arrives too late. Prediction Horizon Should Follow the Control Horizon This suggests that there may not be one optimal prediction length. A warehouse simulator may need long sequences. A robot balancing an object may need extremely fast near-term prediction. A mobile manipulator handling a multi-stage task may need both:
• A long-term plan.
• Short-horizon physical prediction.
• Rapid closed-loop correction. The right design may therefore involve multiple horizons rather than one universally long world model. Public Systems Already Reflect These Different Priorities The differences expressed on the panel are visible in the companies’ public technical systems. ACE Robotics: Unified Understanding, Prediction, and Action At WAIC 2026, ACE Robotics released Kairos 3.1, an action-oriented embodied world model. Public launch materials describe a hybrid Transformer architecture intended to connect:
• Understanding.
• World rollout.
• Action.
• Reflection. ACE Robotics also reported a 125 ms inference latency for an 8B Kairos 3.1 model on NVIDIA Jetson Thor at BF16 precision. That specification matters because the company explicitly frames the model as an edge-capable system for robot action rather than only an offline video generator. Ant Group / Robbyant: From Interactive World Simulation to Embodied-Native Models Robbyant’s open-source LingBot-World began from the generative-video direction. Its official repository describes:
• High-fidelity interactive environments.
• Long-term temporal consistency.
• Real-time user control.
• Sub-second interaction latency.
• 16 FPS real-time generation in the Fast version. Ant Group later integrated LingBot-World-Fast into the LingGuang mobile product, allowing users to convert an image into an explorable generated world. At the same time, Robbyant has expanded into VLA and embodied-model research, making it a useful example of a team that is not limited to one world-model formulation. GigaAI: Pixel Generation and the GigaWorld Platform GigaAI describes itself as an embodied-AI and general robotics company built around three layers:
• GigaWorld, its world-model platform.
• GigaBrain, its general embodied intelligence system.
• Maker, its robot embodiment. The company’s world-model work emphasizes the role of generated physical environments in accelerating data creation, training, and testing. This makes pixel-space generation particularly valuable when the world model is used as a visually rich simulator. AgiBot: World Models as Generative Simulation AgiBot’s public work makes the simulator use case especially explicit. AgiBot Digital World combines:
• 3D assets.
• Physics simulation.
• Expert-trajectory generation.
• Domain randomization.
• Data generation.
• Model evaluation. Its later EVAC framework generates future visual states conditioned on robot action sequences. For this type of application, temporal consistency and accurate reproduction of action–environment interaction are central. X Square Robot: World Model and VLA Convergence X Square’s official company history says its WALL-A architecture deeply integrates a VLA model with a world model. Its 2026 WALL-B direction moves further toward what the company calls a world-unified model architecture. The technical goal is to avoid a pipeline in which a world model predicts independently and a separate policy later translates that result into action. The prediction and control process are instead designed as one end-to-end system. Wujie Dongli: Embodied-Native World Models and Latent Prediction Public information about Wujie Dongli describes the company as building a robot “general brain” and an embodied-native multimodal foundation-model system based on world models. Xia Zhongpu joined the company in March 2026 after leading end-to-end autonomous-driving work at Li Auto. The company’s panel position emphasized latent-space prediction, physical quantities, and JEPA-style thinking. Because detailed public technical documentation for its current world model is still limited, claims about the exact architecture should be treated as the technical direction described by its leadership rather than a fully documented open model. What Is Most Likely to Converge First? The next question was not which architecture would ultimately win. It was: What part of the field will reach agreement first? The answers were very different. Candidate 1: A Unified Representation Ren Guanghui and Xia Zhongpu both emphasized representation. Large language models eventually converged around tokenized sequence representations that made different tasks compatible with one architecture. Embodied intelligence still lacks an equally universal unit. A robot works across:
• Images.
• Video.
• Language.
• Joint positions.
• Velocities.
• Forces.
• Tactile signals.
• 3D geometry.
• Actions.
• Rewards.
• Environment state. If these modalities remain isolated, every system needs complex bridges between them. A unified representation could allow the model to reason across perception, state, prediction, and action inside one common space. The challenge is that physical signals are not naturally interchangeable. A tactile impulse, a camera frame, and a gripper command operate at different rates and carry different information. The “robot token” may therefore be much harder to define than a text token. Candidate 2: Hybrid Architectures Wang Hao’s prediction was architectural fusion. From this perspective, different world-model approaches each solve part of the problem:
• Video generation is strong at visual future prediction.
• Latent prediction can be efficient for reasoning.
• Explicit 3D models can provide spatial memory.
• VLA models connect perception to action.
• Classical control can provide deterministic low-level stability. The final architecture may combine these mechanisms rather than selecting one and discarding the others. This pattern is common in AI. Techniques that initially compete often become components of a larger system after their strengths and weaknesses become clear. Candidate 3: Tactile Sensing Shen Yujun argued that tactile information may become one of the earliest areas of industry agreement. Most large world models have been trained primarily from visual and language data. Robots operate through contact. They need to detect:
• Whether an object is slipping.
• Whether a surface is hard or soft.
• Whether a grasp is secure.
• How much force is being applied.
• Whether a deformable object has changed shape.
• Whether contact has occurred even when vision is blocked. For manipulation, vision can be ambiguous precisely when touch becomes most useful. This creates a strong case for tactile sensing as a native modality rather than an optional sensor added after model training. Robbyant’s research direction also illustrates the broader shift from models derived from digital video toward models designed around embodied control. The important distinction is: A video model is optimized to generate a plausible future.

An embodied world model is optimized to support a successful action.
Those objectives overlap, but they are not identical.
Candidate 4: A Shared Benchmark
Tao Dacheng offered a different answer: convergence may happen first in evaluation, not architecture.
This has historical precedent.
Computer vision did not converge because every researcher agreed on one representation in advance.
Shared datasets and benchmarks such as ImageNet gave competing systems a common measurement target.
Once everyone is measured on the same scale, weak ideas disappear faster and useful ideas spread across architectural boundaries.
That was the logic behind the launch of PHYSICAL IQ at the forum.
PHYSICAL IQ: A Common Yardstick for Embodied Intelligence
PHYSICAL IQ was introduced at WAIC 2026 as a unified benchmark for embodied physical intelligence.
Public launch information says the initiative was jointly started by:

• Shanghai Artificial Intelligence Association.
• Shenzhen Loop Area Institute.
• East China branch of the China Academy of Information and Communications Technology. More than 20 universities and industry organizations were involved. The platform is intended to compare systems across:
• Different robot embodiments.
• Multiple real-world scenarios.
• Different task categories.
• A common evaluation protocol. A useful physical-intelligence benchmark cannot rely only on whether generated video looks realistic. It needs to ask questions such as:
• Did the robot complete the task?
• Was the action physically valid?
• Did it recover from unexpected events?
• How much intervention was required?
• Was the behavior safe?
• Did the policy transfer to a new environment?
• How well did it generalize across objects and embodiments?
• How much time and compute did each decision require? The exact PHYSICAL IQ methodology will need continued public documentation and adoption before it can become a true industry standard. Still, the direction addresses a real problem. World-model claims are currently difficult to compare because one company may report video quality, another reports task success, another reports physical-state error, and another reports simulator consistency. Without a shared yardstick, each system can appear strongest on the metric it was designed to optimize. Visual Realism Is Not the Same as Physical Understanding The debate around PHYSICAL IQ also resembles a broader research question in generative video. Google DeepMind researchers introduced a separate benchmark called Physics-IQ to evaluate whether video models actually understand physical principles. That work found that visual realism and physical correctness can diverge. A generated video may look convincing while violating:
• Solid mechanics.
• Fluid behavior.
• Optics.
• Thermodynamics.
• Magnetism. This distinction matters for robot world models. A visually plausible predicted future is useful only if it preserves the properties required for action. At the same time, a robot does not necessarily need to solve a full physics textbook before moving. The practical standard sits between two extremes: Not enough: Looks realistic but predicts the wrong consequence.

Possibly excessive:
Computes every physical variable even when the action does not require it.

Useful:
Predicts the physical differences that change the robot's decision.
That middle ground is where much of the current research disagreement lives.
From Demo to Deployment
Despite their architectural disagreements, the panelists became much more aligned when the discussion moved from models to real robots.
The final standard is simple:
Can the robot complete the task reliably in the physical world?
Different speakers described this in different language.

• Xia Zhongpu emphasized the effectiveness of the final decision.
• Ren Guanghui focused on how much the world model improves the downstream policy.
• Zhu Zheng emphasized multi-task success rates on physical robots. The terms differ, but the operational question is the same. A world model matters only if it improves the robot’s real behavior. Why a 99% Demo Can Still Be a Bad Product A demonstration can be optimized almost without limit. The team can:
• Reset the environment.
• Choose favorable objects.
• Predefine the route.
• Repeat failed attempts.
• Tune one scenario for weeks.
• Keep an engineer nearby.
• Remove rare edge cases. Deployment removes those protections. A supermarket, hotel, warehouse, or home contains unpredictable events every day. Wang Hao described the cost curve as strongly nonlinear. Improving a system from: 90% → 99% is not necessarily comparable in effort to: 99% → 99.9% The remaining failures are often the difficult long-tail cases. A 1% failure rate may sound small in a laboratory. If a robot performs 10,000 actions per day, 1% can mean 100 failures. Even 99.9% reliability can still produce repeated intervention when the task volume becomes large enough. The business impact includes:
• Human takeover.
• Rework.
• Downtime.
• Customer complaints.
• Safety risk.
• Maintenance.
• Lost throughput.
• Additional supervision staff. This is why deployment changes the meaning of model quality. Random Events Are Normal in the Real World Shen Yujun used a retail example. A robot may be asked to find a package of vegetables. The product is normally stored in one place, but a customer may pick it up and leave it in another refrigerator. From the perspective of a curated dataset, that looks like an unusual exception. From the perspective of a robot serving thousands of customers, exceptions become ordinary. The world model therefore needs to handle:
• Objects in unexpected locations.
• Missing inventory.
• People blocking paths.
• Sudden rearrangement.
• Partial observations.
• Changing lighting.
• Novel object combinations.
• Incorrect assumptions from earlier in the task. This is where generalization becomes more important than replaying a memorized demonstration. Edge Inference Is a Deployment Requirement Tao Dacheng emphasized another practical constraint: latency. Many advanced models run in the cloud. Robots do not always have time to wait for a cloud round trip. A decision window may be measured in tens or hundreds of milliseconds. The network may also be:
• Congested.
• Unavailable.
• Unreliable.
• Too expensive.
• Restricted by privacy or safety requirements. This is why ACE Robotics highlighted the edge performance of Kairos 3.1. A model that is slightly less capable but responds within the control deadline can be more useful than a stronger model that arrives after the action window has passed. For deployment, intelligence must become:
• Cheap enough.
• Fast enough.
• Reliable enough.
• Local enough.
• Predictable enough. This leads to a useful distinction: A demo shows the upper bound of capability. Deployment exposes the lower bound. What Will Look Correct Two Years From Now? The final part of the discussion asked the speakers to imagine looking back from 2028. Which investments made in 2026 will have been right? Which ones may look misguided? The answers again reveal what each team believes is durable. Model Capability Versus Visible Model Output Shen Yujun separated two ideas: Underlying Capability Examples include:
• Generalization efficiency.
• Interactivity.
• Representation quality.
• Action conditioning.
• Data efficiency.
• Causal consistency. Visible Output Examples include:
• A visually impressive generated video.
• A polished demo.
• A single successful long-horizon task.
• High aesthetic quality. A model may improve in its underlying capability without immediately producing the most impressive demonstration. Conversely, a system can be optimized heavily around visible outputs without improving the properties required for general physical intelligence. This is one reason the industry needs better evaluation. Data Infrastructure Is Likely to Remain Valuable Shen also argued that work invested in the data pipeline is likely to remain useful. Even if architectures change, embodied AI will still need:
• Real-world trajectories.
• Tactile data.
• Failure cases.
• Human corrections.
• Robot actions.
• Sensor synchronization.
• Simulation data.
• Quality filtering.
• Evaluation datasets. The uncertainty is not whether data matters. It is whether the data being collected today will still match the representations and training methods used in future models. A dataset can become less useful when:
• Sensor configurations change.
• The action space changes.
• Labels encode outdated assumptions.
• Data lacks necessary modalities.
• The collection policy is too narrow.
• The new model needs a different temporal resolution. Building a flexible data pipeline may therefore be more durable than optimizing one frozen dataset. Do Not Distract the Foundation Model Too Early Zhu Zheng warned that the industry may be exploring too many commercial scenarios before the foundation model itself has stabilized. This is a familiar startup tension. Commercial deployment provides:
• Revenue.
• Data.
• Customer feedback.
• Real constraints. But it can also pull the model team into:
• One-off integrations.
• Custom workflows.
• Customer-specific rules.
• Hardware adaptation.
• Support and operations. If the underlying architecture is still changing rapidly, too much early specialization can slow fundamental research. The balance between model research and commercial deployment will differ by company. There is no universal answer. Embodied-Native Architectures Will Likely Grow Ren Guanghui expects world models to become increasingly embodied-native. That implies models designed from the start around physical interaction rather than models adapted later from internet video or image generation. Embodied-native training may include:
• Robot action.
• Proprioception.
• Force.
• Tactile data.
• Egocentric video.
• Multi-view observation.
• 3D state.
• Tool feedback.
• Intervention data.
• Task success and failure. The architecture may also be designed around the control loop itself. The important question is whether the model represents the variables a robot needs to act, not whether it resembles a media-generation model internally. The Most Interesting Opportunity Is the Disagreement The panel did not end with a single architecture everyone accepted. That is probably a sign of a young field rather than a broken one. Several important areas remain unresolved: Question Competing views What should the model predict? Pixels, latent state, explicit geometry, or a hybrid How accurate must physics be? Exact state estimation vs. action-relevant plausibility How far should it predict? Long-horizon consistency vs. low-latency short-horizon control What will unify the field? Representation, architecture, tactile data, or benchmarks Where should inference run? Cloud, edge, or hybrid What is the real success metric? Simulation quality, physical prediction, policy gain, or real task success What data matters most? Video, robot trajectories, force, touch, simulation, or a mixture In mature fields, architectures often look more similar. In emerging fields, disagreement can reveal where the largest unsolved opportunities are. A team that proves one of these currently disputed assumptions may create a technical advantage before the rest of the industry converges. A Practical Framework for Evaluating a Robot World Model For developers and robotics teams, the panel debate can be translated into a more concrete evaluation checklist.
• Prediction Target Ask what the model predicts:
• Pixels.
• Latent state.
• Geometry.
• Action.
• Force.
• Future observation.
• A combination of these. The output should match the downstream control problem.
• Prediction Horizon Measure performance at:
• Immediate next step.
• Short horizon.
• Multi-second horizon.
• Long task horizon. Do not evaluate only the horizon that makes the model look strongest.
• Latency Measure actual inference time on the intended deployment hardware. A model that misses the control deadline should not receive full credit for accurate prediction.
• Physical Validity Test:
• Collisions.
• Contact.
• Balance.
• Deformation.
• Friction-sensitive actions.
• Object permanence.
• Multi-view consistency.
• Policy Improvement The most useful question may be: Does adding the world model improve real task success? Compare the downstream policy with and without the world-model component.
• Generalization Test new:
• Objects.
• Layouts.
• Robot bodies.
• Lighting.
• Materials.
• Human behavior.
• Task combinations.
• Recovery Measure what happens after the first mistake. A deployment system should be able to detect failure, update its state, and try another plan.
• Edge and Cloud Behavior Test both normal and degraded networking. A robot should fail safely when remote inference is unavailable.
• Intervention Rate Track human assistance per:
• Hour.
• Task.
• 100 actions.
• 1,000 actions. This often reveals deployment quality more clearly than one headline success rate.
• Cost per Successful Task Include:
• Model inference.
• Hardware.
• Network.
• Human intervention.
• Rework.
• Maintenance.
• Energy.
• Downtime. The best world model is not necessarily the one with the highest benchmark score. It is the one that makes the complete robotic system work better. FAQ What is a world model in embodied AI? A world model is a model that represents or predicts how the environment may change in response to time, robot actions, or other events. In robotics, it can support planning, simulation, action selection, training-data generation, or policy improvement. What were the main world-model approaches discussed at WAIC 2026? The discussion broadly covered pixel-space generation, latent-space prediction such as JEPA-style approaches, and hybrid or unified systems that combine world prediction with VLA control, 3D reasoning, or other representations. The categories overlap because several companies use multiple techniques. Does a robot world model need an exact physics simulator? Not necessarily. Some researchers argue for accurate prediction of geometry, motion, and force, while others prioritize physical reasoning that is accurate enough to choose the correct action. The required precision depends on the task and cost of failure. Why is long-horizon consistency controversial? Long rollouts are useful for simulation and long-term planning, but they increase inference cost and may slow real-time control. A physical robot may need short, fast predictions for immediate action while using another mechanism for long-term planning. What is PHYSICAL IQ? PHYSICAL IQ is an embodied physical-intelligence benchmark initiative launched during WAIC 2026. It is intended to provide a common evaluation protocol across different robot bodies, scenarios, and tasks. Is PHYSICAL IQ the same as Google DeepMind’s Physics-IQ? No. They are separate projects. PHYSICAL IQ is the embodied-robotics evaluation platform introduced at WAIC 2026, while Physics-IQ is a research benchmark for testing whether generative video models understand physical principles. Why is tactile sensing important for world models? Vision cannot fully reveal contact force, slip, pressure, compliance, or some material properties. Tactile and force signals give robots direct information about physical interaction and may become increasingly important for manipulation-focused world models. What matters more than a good robot demo? Deployment reliability. Teams need to measure real task success, latency, intervention rate, recovery from unexpected events, safety, cost, and performance across many environments rather than only selected demonstrations.