NVIDIA, Baidu, Unitree... 32 World Models Unveil the Hidden Trumps of Physical AI
To make a robot pick up a cup, it has to solve far more problems than just identifying the object: where the cup is, from which direction the robotic arm should extend, when the gripper should close, and whether it will hit the edge of the table during movement. When an autonomous vehicle passes through an intersection, it also needs to make judgments: if it accelerates now, how will the nearby vehicles and pedestrians act next.
This type of AI that requires perception, decision-making and action in the real physical environment is commonly referred to as physical AI. Humanoid robots, autonomous vehicles, robotic arms and drones are all its application carriers. Weight, inertia, friction and collision will directly determine whether they can successfully complete a task.
The world model is a type of model that helps AI predict environmental changes, which can be understood as a data-trained simulator inside the system. It receives current observations, combines conditions such as actions, deduces what may happen next, and provides a basis for planning, training and evaluation.
Taking picking up a cup as an example, the model can try to predict: if the robotic arm moves along a certain trajectory, whether it can approach the cup; after the gripper lifts, whether the cup will move accordingly, and if it is likely to slip. By sending the future results corresponding to candidate actions to the planning system for comparison, inappropriate solutions can be identified before actual execution.
The output forms of these predictions are diverse. Some models generate future videos, some construct explorable 3D scenes, and some represent environmental changes in compressed abstract features. Accordingly, researchers will examine the consistency between images and space, changes caused by actions, and whether the predictions can help the system make better decisions.
World models can also provide training experience. Robots can conduct repeated trials in the environment generated by the model, and researchers can also generate task data that is difficult to collect, or evaluate strategies before deployment. Some teams further jointly train future prediction and action generation, so that the model can simultaneously learn how the environment changes and how the robot acts.
These capabilities still have specific limitations: objects may deform or disappear during interaction, collisions and slips may be inaccurately simulated, and continuous prediction will also accumulate errors. Therefore, the application value of an achievement should be evaluated in combination with its contribution to real tasks, applicable scope and operation cost.
On November 3, 2026, Baidu first unveiled Apollo World, enabling autonomous vehicles to complete continuous trips in a virtual city; a week ago, Google DeepMind released Genie 4, where users can change the weather or add characters while exploring the environment. Research on world models is being carried out in directions such as robot control, autonomous driving simulation and interactive environment generation.
The following 32 models and research systems are selected based on comprehensive technical representativeness, public experiments, application progress and ecological impact, including representative early-stage studies. The serial numbers are for listing purposes only and do not represent cross-task performance rankings. All information is as of November 4, 2026.
01 NVIDIA Cosmos 3: Integrating world prediction, reasoning and action generation into one system
On June 1, 2026, NVIDIA released Cosmos 3. This generation of models adopts a hybrid Transformer architecture, which combines visual reasoning, world generation and action prediction, and can process and generate text, images, videos, environmental sounds and action information.
Its target customers include developers of robotics, autonomous driving and industrial vision: it can not only predict how the environment will change next, but also generate training and evaluation data, and serve as the foundation of the world action model. NVIDIA has provided versions such as Super and Nano, and opened model resources. The Cosmos layout covers models, data processing and deployment tools, allowing developers to complete multiple physical AI tasks around the same system.
02 Google DeepMind Genie 4: An interactive world that can evolve along with operations
On October 27, 2026, Google DeepMind released Genie 4. Users can move the perspective, change the weather, add characters and objects during the generation process, and the model will continue to deduce scene changes, and simultaneously generate environmental sounds, music and dialogues.
Genie 4 introduces the capabilities of the Gemini series to improve the continuity of images, behaviors and scenes. Its uses include training environments for robots and agents, as well as interactive content creation. At the time of release, the model was open to trusted testers in the form of research preview. It demonstrates the ability of continuous generation and real-time environment intervention, while the robot training effect still needs to be evaluated in specific tasks.
03 Meta V-JEPA 2 / 2.1: Predicting the physical world in abstract features
Meta released V-JEPA 2 in June 2025, and further released V-JEPA 2.1 in March 2026. This series encodes videos into abstract features, learns object motion and scene changes, and then uses prediction results to assist planning, reducing the dependence on pixel-by-pixel image generation.
V-JEPA 2 is pre-trained with data including more than 1 million hours of videos, then added with robot action data to carry out grasping and placing experiments. Version 2.1 focuses on improving local spatial features, depth understanding and temporal consistency, and is used for robotic arm planning and navigation research. The code and models have been made public, representing the world model route of predictive visual representation.
04 World Labs Atlas: Connecting images, videos and 3D space
On September 1, 2026, World Labs released Atlas. The model natively processes text, images, videos and 3D information, organizes inputs into shared spatial context, and is used to generate new perspectives, reconstruct scenes and simulate spatiotemporal changes.
Atlas can generate images seen at the specified camera position based on one or more pictures, and can also output explicit 3D results. The company demonstrated camera-controlled videos of up to about one minute at 1440p, as well as a real scene-to-simulation process for robots. Atlas will support subsequent products such as Marble, and early experience applications were open at release. Its main focus is on spatial generation, reconstruction and simulation scene construction.
05 Odyssey-3: Extending from world simulation to robotic arm and vehicle experiments
Odyssey released Odyssey-3 on October 8, 2026. The model adopts an autoregressive diffusion Transformer, and predicts how objects move, interact with each other, and how the scene will continue to evolve based on previous observations and the latest inputs.
In addition to the real-time interactive environment, the company also demonstrated experiments on robotic arm operation, humanoid robot and vehicle control. For example, with the backbone of the world model frozen, 20 hours of driving data was used to train the driving strategy, and closed-loop experiments were carried out on real roads in India. These results demonstrate the path for world models to adapt to different hardware. Odyssey-3 is currently available for research preview, and its related capabilities come from the experiments disclosed by the company.
06 Runway GWM series and Praxis-1: Extending from interactive simulation to robot operation
Runway's GWM-1 family includes three directions: explorable environments, virtual characters and robot operations. Among them, GWM Robotics generates future interaction images based on robot actions, supports multi-view and long context, and is used for strategy evaluation, synthetic data augmentation and the visual foundation of robot strategies.
In September 2026, the company further announced the world action model Praxis-1, which applies video pre-training to real robot control. At the time of release, the model was being tested with early partners, and public weights are still in the follow-up plan. GWM Worlds 2, released in the same month, is oriented to real-time audio-visual interactive environments, providing continuous generation at 720p and 24 frames per second.
07 Waymo World Model: Generating the world seen by both cameras and LiDAR
On February 6, 2026, Waymo publicly released the Waymo World Model. It is based on Genie 3, adapted for autonomous driving scenarios, and can generate camera images and LiDAR data at the same time.
Engineers can adjust driving actions, road layouts and text instructions to simulate rare scenarios such as oncoming vehicles driving against the traffic, water accumulation, and extreme weather. They can also start from the same real record to test what will happen if another driving route is taken. This model serves Waymo's autonomous driving simulation system, focusing on expanding test scenarios that are difficult to cover through road collection.
08 Wayve GAIA-4: Allowing the decisions of the driving model to change subsequent simulations
Wayve released GAIA-4 on August 3, 2026. It builds a simulation starting point from real sensor records, and connects the driving model to a closed loop: after the vehicle chooses to brake or steer, the subsequently received perspective and sensor information will change accordingly.
GAIA-4 supports an evaluation method that retains the original trajectories of other traffic participants, which is convenient for reproducing safety-critical events; it can also enable traffic participants that react to the ego vehicle's behavior to study interactive scenarios such as lane merging. Wayve uses it to analyze safety operator takeover events, retest model updates, and compare the consequences of different driving decisions in the same scenario.
09 Waabi Copilot4D: Directly predicting the next step of the 3D point cloud
Copilot4D is Waabi's public 4D world model research, published at ICLR 2024. It converts LiDAR point clouds into discrete representations, predicts future observations through a diffusion model, and learns the changes of road environments in 3D space and time.
The model can take candidate driving actions as conditions to deduce the environment after different actions are taken. For example, when the ego vehicle brakes suddenly, how the following vehicles may respond. The team conducted tests on datasets such as nuScenes, KITTI and Argoverse 2. This research provides a path to learn physical dynamics from sensor data, and also enables autonomous driving world modeling to cover observation forms beyond videos.
10 1XWM: Generating task videos first, then converting them into humanoid robot actions
1X publicly disclosed 1XWM integrated into the NEO robot in January 2026. The system receives text instructions and starting frames, generates future videos of task execution, and then converts the frame changes into robot action trajectories through an inverse dynamics model.
Its training combines Internet videos, first-person human data and NEO data. In the version disclosed in January, it takes about 11 seconds for the world model to generate a 5-second video, and action extraction takes another about 1 second. Response speed is still the key point for improvement. In June 2026, 1X announced the establishment of the World Model Lab to continue expanding the learning process between pre-training, robot deployment and data recycling.
11 NVIDIA DreamDojo: Training robot simulators with large-scale human videos
DreamDojo, disclosed in February 2026, is a general world model oriented to robot interaction prediction. The research is based on Cosmos-Predict2.5, pre-trained with about 44,700 hours of human interaction videos, and then adapted through robot data.
Given robot actions, DreamDojo predicts corresponding environmental changes. The team improves the inference speed through distillation, and uses it for interactive teleoperation, strategy evaluation and model-based planning. Compared with general scene generation, it pays more attention to whether actions can cause reasonable object changes, and whether it can handle objects and interactions not covered in the training data of the target robot. The project has made public code and model resources.
12 Ant Lingbo LingBot-World 2.0: Extending the interaction length of the real-time world
LingBot-World 2.0, also known as LingBot-World-Infinity, released its technical report, inference code and models in July 2026, and supplemented more model versions in September. It improves the generation quality and response speed during continuous interaction through causal pre-training and distillation.
The official demonstration configuration supports 720p, 60fps video stream, with interactions covering movement, more character actions and text-driven events, and also introduces agents responsible for character behavior and environmental changes respectively. Its main achievements focus on sustainably explorable interactive environments. The project opens multiple models of different scales and types, and the current license is limited to non-commercial use.
13 Tencent Hunyuan HY-World: Generating savable and editable 3D scenes
Tencent publicly disclosed HY-World 2.0 in April 2026, then released code and models in stages, and announced the HY-World 2.1 product update in July. This series supports inputs such as text, single images, multi-view images and videos to generate or reconstruct explorable 3D scenes.
Version 2.0 organizes panorama generation, trajectory planning, world expansion and 3D reconstruction into a complete process, and outputs assets such as 3D Gaussians and meshes, which can be connected to game engines and tools such as Isaac Sim. It focuses on the generation and reconstruction of environmental spaces, and can provide scene materials for robot simulation; dynamic processes such as object force and contact still require corresponding physical modeling support.
14 Agibot Genie Envisioner 2.0: Allowing robot strategies to conduct continuous experiments in the model
The core achievement of Genie Envisioner 2.0, GE-Sim 2.0, is a closed-loop video world simulator for robot operation. The project was launched in April 2026, the technical report was made public in May, code and models were released in June, and evaluation support continued to be updated in October.
The model is trained with data covering teleoperation, robot strategy deployment and object interaction. After the external strategy gives an action, the simulator generates a new environmental observation for the strategy to continue decision-making. Public examples include the success and failure processes of tasks from head and dual-wrist perspectives. This system connects the world model to the strategy evaluation and training process, and supports access to external reward models.
15 Motus Tech Motus2: Connecting actions, predictions and evaluations in the same model
Motus2, disclosed in 2026, is oriented to dexterous operation, and uses the same set of shared parameters to provide three interfaces: the strategy proposes candidate actions, the simulator predicts the visual changes caused by the actions, and the evaluator judges the results, and then feeds back to the strategy for improvement.
The original recorded scale of human data disclosed by the project is about 130,000 hours. The training expands from monocular first-person videos to synchronized binocular data, and then adapts to robot trajectories. It also studies long-term memory and haptic feedback, and conducts experiments on platforms equipped with hardware such as dual arms and dual dexterous hands. The project has published papers, demonstrations and phased open source plans.
16 Manifold Space WorldScape: Unified modeling of spatial movement and object operation
Manifold Space released WorldScape in February 2026, which integrates spatial navigation and object operation into the same interactive generation process. The model introduces geometric constraints and memory mechanisms to improve problems such as scene structure drift after continuous movement and appearance changes of historical objects.
In July 2026, the company further announced WorldScape v2 and Policy v2, expanded different action control methods, and added memory and contextual learning capabilities for long-term robot tasks; in October, it announced R²SI robot self-improvement research. Its product research involves both interactive world generation and world action models, trying to connect environmental prediction, task execution and experience utilization.
17 AutoNavi DreamX-Phi 1.0: Predicting interactive images around robot actions
DreamX-Phi 1.0 is a robot operation world model disclosed in 2026, built based on Wan2.2-TI2V-5B. After inputting the initial observation and the specified robot action sequence, the model generates corresponding future videos, focusing on maintaining the consistency of robot movement, scene geometry and the state of the operated object.
It adds signals such as the pose of the end of the robotic arm and the state of the gripper to the action conditions, uses interaction-related training to improve object changes, and improves inference efficiency through few-step generation. The team reported the results of the relevant track in WorldArena 2.0. The official repository has published technical descriptions, and the weights and inference codes are listed as content to be opened after the competition ends.
18 Ctrl-World: Evaluating and improving robot strategies in "imagination"
Ctrl-World was publicly disclosed in October 2025 and selected for ICLR 2026. The model jointly predicts the images from the external camera and the wrist camera, controls future changes through frame-by-frame action conditions, and retrieves historical memory to maintain the consistency of multi-step interaction.
The team trained with DROID data, compared the simulation of the model with the performance of real robots in tasks such as grasping, folding towels, opening and closing drawers, and used the screened synthetic successful trajectories for strategy fine-tuning. The paper reports that the success rate on specific new instruction and new object tasks increased from 38.7% to 83.4%; the model still has errors in simulating fine dynamics such as collision and sliding.
19 Dreamer 4: Allowing agents to train behaviors in the learned world
Dreamer 4, publicly disclosed in September 2025, combines the