Physical AI shifts the focus of robotics from rigid, hard-coded automation to an entirely new realm, in which machines can survey the real world, perceive their environments, reason about it and take actions without human intervention. It’s a bold new frontier, and it’s going to require massive amounts of data to make it happen – far more, in fact, than what is available today.
But AI gives us the tools to overcome this notorious “sim-to-real” gap. With the rise of AI “world models,” it’s possible to create simulated environments that replicate the laws of physics and spatial relationships, to generate all of the training models that robots will ever need.
This is a different approach from Vision-Language-Action (VLA) models, which learn to map perception and instructions directly to robot actions. Instead, world models simulate how an environment behaves, generating the training data and mental rehearsal space a robot’s policy (VLA or otherwise) can then learn from.
World models are such a nascent concept that current research efforts have explored their development in remarkably different directions. Depending on whom you ask, the best approach might involve mapping geometric 3D spaces, building photorealistic simulators or abandoning visual generation entirely to focus on more abstract forms of cognition.
Let’s take a look at how five of the top AI research labs are building world models for the explicit purpose of training intelligent robots.
1. Google DeepMind’s Genie 3
Described as a pure-play generative pixel engine, Genie 3 is the most promising world model developed by Google’s DeepMind research unit. Traditional game engines, from which world models evolved, rely on hard-coded physics, but Genie 3 abandons this in favor of a neural physics engine that’s designed to generate interactive and explorable 3D environments on a frame-by-frame basis.
It’s based on the classic autoregressive transformer architecture found in every LLM, only it has been adapted to generate dynamic visual sequences rather than text. Instead of ingesting words, it devours raw video data before spitting out pixels as discrete visual tokens. DeepMind’s Latent Action Model, a technique carried over from earlier Genie versions, lets Genie 3 infer the actions it sees in videos without the need for labels.
Attention mechanisms are used to predict the next frame, and it does this at 720p resolution and 24 frames-per-second. Every frame is conditioned on the actions in the preceding frame. It also supports “text-promptable events,” enabling developers to introduce phenomena such as new weather conditions or environment rules at any moment.
The advantage of Genie 3 is its ability to generate visually striking and highly interactive robotic training playgrounds with unlimited variations. By pairing it with agents such as SIMA 2, Google says it can teach robots to follow text-based prompts and adapt to changing environments on the fly.
Google says Genie 3 was purpose-built for generating simulated playgrounds for robot training. With its low latency and its geometric awareness enabling spatial consistency, it’s well-suited to robotic workflows and can generate multi-view outputs. Being action-conditioned means it can support closed-loop training scenarios, and its API access allows direct integration with robotic development pipelines.
Genie 3 enables AI agents to navigate realistic simulations that conform to physics, so that robot models can learn the spatial reasoning necessary to operate in the real world.
2. Decart’s Oasis 3
Decart says Oasis 3 is purpose-built to overcome the bottlenecks of robotic reinforcement learning, fusing generative photorealism with real-world physics to enable low-latency closed-loop simulations in real time.
Oasis 3 is based on a Diffusion Transformers architecture that sits atop the proprietary Decart Optimization Stack (DOS). It was co-designed using Nvidia’s physical AI ecosystem and is capable of generating interactive environments with realistic physics properties at 512x768x3 resolution. Its native three-camera view is key to maintaining consistent spatial understanding and temporal coherence from multiple viewpoints, so robots can accurately judge characteristics such as textures, depth and peripheral environments.
Like Google’s Genie 3, Oasis 3 is available through an API, but the difference is it sits directly inside live robotic training loops. This means developers can adapt their simulations on the fly with natural language prompts and create rare edge cases to train robots for any situation. With autonomous vehicles, for example, developers can introduce interactive hazards such as other cars driving into oncoming traffic lanes, flooding, tricky visibility scenarios, and other phenomena to teach their systems how to react.
Previously, automated systems could only be trained by recreating these scenarios in real-world training environments, but this method is expensive to scale. Oasis 3 is far more cost effective as teams can generate infinite simulations to teach models how to deal with every conceivable risk. It can be applied to other kinds of physical AI use cases, including humanoid robots, autonomous boats, unmanned drones and more.
3. World Labs’ Marble
The brainchild of AI pioneer Fei-Fei Li, World Labs is focused on spatial intelligence, rejecting the traditional idea of flat 2D video streams. The company aims to teach robot AI models how to understand, reason and build within persistent 3D geometries.
Its flagship model Marble uses abstract techniques such as 3D Gaussian Splatting, which represents scenes as dense clusters of semi-transparent particles to enhance visual fidelity. It isolates its static environment layouts from moving elements, generating a persistent 3D background paired with a “collider mesh” for mapping depth and physical boundaries.
According to World Labs, Marble can output extremely high-resolution, geometrically-aware 3D simulations of almost any environment. It’s extremely flexible, with users able to rotate the environment, adjusting the camera view to any perspective. The structural collision mesh can be plugged into traditional physics engines such as Nvidia’s Isaac Sim to enable tightly-controlled, semantic modeling with high-accuracy physics.
The downside to this approach is that Marble’s environments are intrinsically static, for they lack object dynamics. To integrate moving elements, such as a robotic arm or a vehicle, these must be imported using external platforms such as RoboSuite. Nonetheless, the infinite structural variation combined with its persistent representations of space, volume and depth can be game-changing for generating accurate robot training data.
4. Meta’s V-JEPA 2
Yann LeCun, who served as Meta’s Chief AI Scientist at the time, designed the Joint Embedding Predictive Architecture, or JEPA, framework in an effort to improve upon inefficient media generation techniques. The framework, first announced back in 2022 and rolled out a year later, intentionally avoids pixel generation in favor of predicting abstract concepts.
Instead of autoregression or diffusion, JEPA consists of a Context Encoder that translates partial visible data into compressed mathematical representations; a Target Encoder for processing the complete data into target embeddings; and a Predictor that infers the target embeddings of hidden sections based on the context embedding.
A variation of JEPA, called LeJEPA, introduced random 1-D projections to prevent the mathematical representations from degrading during long training sessions.
LeCun has argued that V-JEPA 2 offers several advantages for robotics training, with one of the main ones being its speed and scalability. The architecture is immune to visual hallucinations because there are no visuals at all. Instead, physical AI models learn via abstract mathematical representations of the physical world.
The lack of video outputs means V-JEPA 2 dramatically enhances compute efficiency, as there’s no resource-intensive graphics generation. This means it can serve as a high-level, energy-efficient cognitive engine for edge robotics, enabling predictive reasoning on less sophisticated processing hardware.
V-JEPA 2 was developed at Meta FAIR, the research lab LeCun founded and led for over a decade. In late 2025, LeCun left Meta to launch AMI Labs (Advanced Machine Intelligence Labs), a Paris-based startup built around the same JEPA-based, non-generative approach to world models, effectively taking his research thesis independent.
Notably, though, AMI Labs has yet to ship a product. Its own leadership has said applied outputs are roughly a year out, with robotics named as one of the first target domains alongside healthcare and industrial automation. That makes V-JEPA 2, built under Meta’s banner, the most concrete, tested expression of LeCun’s world model philosophy currently available, even as the research program’s future center of gravity shifts to his new venture.
5. Verses.ai’s AXIOM
Another non-video world model, Active eXpanding Inference with Object-centric Models (AXIOM) was developed by Verses.ai Chief Scientist Karl Friston, a renowned British neuroscientist and theoretician at University College London.
Interestingly, this model abandons the idea of massive neural networks in favor of first-principles neuroscience. Active inference is paired with variational Bayesian updating to create modular and gradient-free representations of the physical world.
The architecture merges four models, including a slot mixture model that parses scenes and breaks them down into isolated “slots,” such as a person in a scene, the hat they’re wearing, their shoes and so on.
An identity mixture model then applies semantic labels to each slot, while a transition mixture model provides a cognitive library of “motion prototypes” (verbs like falling, sliding or rolling) to predict the slots’ behavior. Finally, there’s a recurrent mixture model that reviews contextual snapshots of the generated environment in real time to determine what should happen next.
AXIOM continuously updates its understanding, balancing goal-oriented actions with information-seeking exploration. Should it stumble across something unexplainable, it updates its internal mathematics and creates new rules to predict what should happen next. It boasts a comprehensive understanding of physics, allowing robots to transfer their knowledge of inertia or gravity to objects they’ve never come across before, so they can predict their actions in real time.
AXIOM excels at providing explainable reasoning to intelligent machines in unpredictable human environments. However, its usefulness is currently limited due to the lack of developer tooling that makes it difficult to integrate into existing robotics training pipelines.
The Future of Physical AI
World models are at the bleeding edge of AI development, which is why AI labs are exploring such vastly different model architectures. Each one has appreciable advantages, with Decart’s and Google’s models leveraging raw generative power to build highly flexible photorealistic simulations and train robots through infinite trial-and-error.
World Labs’ focus on spatial structure has the advantage of enabling stronger 3D awareness for navigation and the manipulation of physical objects. Then there’s Meta and Verses.ai, which both ignore visual generation altogether to focus on internal logic. They’re focused on compact and efficient models that excel in causal reasoning and real-time adaptation.
In future, it’s likely that physical AI won’t be based on any single one of these paradigms, but an amalgamation that leverages the best aspects of them all. With a system that combines Oasis’s real-time simulation capabilities, Marble’s spatial awareness and the cognitive reasoning efficiency of AXIOM, the possibilities for physical AI might be as infinite as the simulations it creates.
