The robot opens the beer bottle lid, the beer bubble bursts, and the liquor flows into the glass from the bottom right corner of the image.
It looks like a normal live-action video. It's hard to imagine, but in reality, robots, objects and even the future state of beer flowing in the image come from the video generated by the model after reading the robot's actions.
The images are from Current Robotics' latest release CurrentWorld-0, their second work after the release of Curr-0.
The team describes it as an interactive world simulator (Interactive World Simulator). It differs from conventional physics engines: it does not rely on artificial writing of physical formulas, but instead learns laws from real-world data to form a data-driven simulation.
Similar trends have been seen in the past, such as 1X's 1X World Model, Ctrl-World in Physical Intelligence and SC3-Eval in WorldGym in DeepMind, and Cosmos Predict in the UK.
But this is the first time that cross-body, multi-point, and force-touch predictions have been integrated into the same world simulation machine while simultaneously simulating, learning, and interacting with human and robotic models.
In this world, what you're seeing is not just simple clawed arms -- mobile robots, two-legged humanoid robots with clever hands that can be modeled and controlled.
The objects they handle aren't just the body. Robots can fold their heads, pick up pillows, wipe bread, or cut cucumbers. Both flexible and fluid objects can stay in a reasonable state. It's not really about how much the image looks like a real video, but how much the model still needs to evolve after the robot's movements change, with different points of view, different objects and contact processes along the same physical event.
Scaling Evaluation：
Robots also need a field that can expand indefinitely
What can a world simulation like this do? A very straightforward answer is to evaluate robots.
Today, we're training a big model and we can run thousands of benchmark cases at the same time, and we know quickly where the model has improved and where it has retreated.
But when it comes to robots, it's all over again. The number of robots, the sites, the man-made care and the replication of tasks, all limit the size of the test. The real world is irreplaceable, but it's hard to be a place where you can replicate it infinitely.
Traditional physical simulation solves some of the problems. Simulators such as Isaac Sim can run a large number of experiments in parallel, but on the assumption that you must first know how to model the world. The rigidity, friction and joints can be artificially defined, but the flexible objects behind the curls, breadcrumbs, rolling beer, the fluid and complex contact can be difficult to write into a physical engine.
Is it possible to immerse the real world itself in an environment that can operate without limits?
This is exactly what CurrentWorld-0 wants to do: learn directly from real interaction about what happens next when a robot makes a move.
If predictions are reliable enough, a large number of strategies can first run, fail and compare in a model-generated world, and then bring the most critical problems back to reality.
But the world model used for robotic evaluation cannot simply produce a realistic video. With one robot, the laws of the world cannot change; with one perspective, only observation can change; with contact, vision, sensation, and touch must describe the same physical process.
As a result, CurrentWorld will focus its capabilities in three directions: Cross-Embodiment, Multi-View, and Force-Tactile.
They end up solving the same problem:
Regardless of who is watching, who is acting, the world should be the same.
The solution to CurrentWorld-0:
What kind of world simulation can actually be used to verify robots?
First of all, there is a most basic premise: that the robot's actions must be real algorithms.
Video models have strong visual and motor precedents, but this is not always a good thing for a robot's evaluation. While the mechanical arm has been scratched, the model may continue to be generated along a more familiar trajectory, allowing the object to rotate automatically with its claws; the task actually fails, but the image completes a successful outcome.
As soon as the world model replaces robots with errors, evaluation also loses its meaning.
Current Robotics, previously published by dWorldEval (published in ICML 2026 Spotlight), studies the controllability of actions: whatever the robot performs, the environment must give corresponding results.
On this basis, CurrentWorld further addresses three issues.
The first is to allow different bodies into the same world.
The freedoms of fixed-arms, mobile-arms and humanoid robots, as well as their kinetic structures and controls, are different, so CurrentWorld does not force them to share a set of action spaces, but instead reserves their own Action Subspace for different robots.
It's not the same as the body, it's the same as the physical laws that they follow when they operate in the world.
The goal of Cross-Embodiment is to allow different robots to operate in their own ways while sharing an understanding of the same world.
Second, let's see the same world from different points of view.
The head, wrist and third-point cameras see different intersections of the same interaction. The object in one perspective moves and the state in the other perspective must change simultaneously; even if briefly obscured, it cannot change a world view when it appears again.
CurrentWorld combines multiple synchronous viewpoints to model the same set of physical states.
The point of view can change, but the state of the world cannot change with the perspective.
Third, let's make the world not just RGB.
Many robotic tasks cannot be judged by vision alone whether the interaction is really successful. It is the same as holding an object, the image may be almost the same, but once it is firmly held, the other is slipping; in tasks such as cutting the skin, opening the bottle, pulling out, the real results are often determined by contact and force.
As a result, CurrentWorld not only predicts future images, but also sensations and sensations.
Vision tells the robot what's going on, force and touch tell it further: whether the contact is actually happening, and how the contact is changing.
This allows CurrentWorld to model not just a visually sound video, but how the entire physical interaction continues to evolve after the robot moves.
Failure is not the end.
It's the beginning of the next round of training.
But it's not enough just to tell you that the robot has failed.
More importantly, the question is: can we continue to start from a failed position and generate new training data?
In CurrentWorld, the strategy can be executed autonomously. When the robot enters a potentially failed state, humans can take over directly and complete corrections by remote operation; the same intermediate state can also be saved, rolled back, and tried different recovery methods.
This is a continuation of Current Robotics' previous idea of Hi-WM (Human-in-the-World-Model, published in the RSS 2026 Robot World Models Workshop): to turn the original one-time failure into a data node that can be explored repeatedly.
In CurrentWorld, human intervention involves not only movement trajectories, but also synchronous visual, sensory and tactile feedback.
Therefore, evaluation is no longer just the end of training, it is becoming the entrance to the next round of training data.
Where robots fail, new experiences come from.
The model of the world.
Finally, we're going back to the whole link of robot learning.
Putting CurrentWorld back on the current robotics technology route will make its location clearer.
HumanEx provides real human behavior data, from which Curr-0 learns full-body mobility and agility; the strategy enters CurrentWorld to receive assessments; exposes failures, and then introduces human takeover near failures through Hi-WM to generate new error-correcting data.
This is a complete cycle:
Real experience → Model learning → World model assessment → Exposure to failure → Human correction → New data → Re-training
The real world has always been the source of experience and the place where robots are ultimately tested. What CurrentWorld has done is to extend the space between two real executions, allowing strategies to be verified more quickly, allowing failures to be revisited, and allowing new experiences to continue to grow out of failure.
Robots don't get smarter by generating more images that are correctly framed.
The real ability to move forward is that every action has consequences, every failure can turn into a new experience and ultimately return to the real world to be tested.