Neo Gamma 2.0 WORLD MODEL AI News SHOCKS Robotics Industry ($30000 HUMANOID)

1x has introduced a new model, the 1x world model, that allows its Neo Robot to predict the results of its actions in a simulated real world. This model can forecast detailed outcomes of different actions, making it more efficient than traditional methods. It improves with more data, enhancing its prediction accuracy, and can transfer knowledge across different tasks, making it versatile. The model’s predictions are highly accurate, often matching real-world results, which could hint at a future where robots learn more holistically.

1x just unveiled its new 1x world model, allowing humanoids to understand the future and how it relates to their actions right now. And that’s only the beginning of what the Neo Gamma Robot can now do as a direct result. So stay tuned for this and more on AI news, starting with explaining what the 1x world model is. Put simply, it’s a predictive engine that lets 1x’s Neo Robot anticipate the outcomes of its actions in a simulation of the real world. For example, from one starting state, the 1x world model can predict four separate futures for Neo when conditioned on four different low-level action sequences.

And these predictions aren’t vague guesses, but instead they’re detailed forecasts of how objects move, how the robot’s body interacts with them, and whether the task succeeds. And it matters because testing every possible action in a real home would take forever, as traditional physics-based simulators require painstakingly crafting virtual environments, while mock real world setups are costly and limited. But the 1x world model bypasses these hurdles and slashes evaluation time while incorporating another key aspect, its action controllability. And this is unlike most video generation models, which instead rely on text prompts like pick up a cup, whereas 1x world model is action controllable.

In fact, it takes exact robot trajectories, which are just specific sequences of motor commands, and predicts their outcomes. For example, given a few initial video frames of a kitchen, 1x world model can actually simulate what happens if Neo opens a door, or if it wipes a cloth across a counter, or picks up a mug. And these simulations capture real world physics like the swing of a door or the drag of a cloth. And this precision lets 1x compare different policies side by side to choose a specific AI strategy for each task.

Like if one policy grabs the mug too forcefully and another does it gently, 1x world model can predict which is likelier to succeed without ever even touching a real mug. And beyond this leap in controllability, there’s another leap in scaling. This is because 1x world model gets smarter as it’s given more data, with 1x testing this by training the model on several tasks, including operating an air fryer, playing an arcade game, and organizing a shelf. And the results clearly demonstrated that more data boosts prediction accuracy. For instance, early versions of the 1x world model struggled with the air fryer task, treating the tray and the body as one object and hallucinating the whole unit sliding off the counter.

But after being given more training data showing how the tray moves within the fryer’s base, the model subsequently mastered the interaction, even capturing subtle constraints like the tray’s limited range. Similarly, adding data from arcade and shelf tasks improved accuracy across all three, showing that the model learns to generalize from one task to another. And this brings us to multitask transfer, which is 1x world model’s ability to transfer knowledge across separate tasks. And this brings us to multitask transfer, which is 1x world model’s ability to transfer knowledge across separate tasks. And the results were shocking, because training on shelf organizing data didn’t just improve shelf predictions, it also made the model much better at arcade tasks, which is totally separate.

In fact, this positive transfer suggests 1x world model isn’t just memorizing tasks, but actually learning the underlying principles, like how objects move or how surfaces interact. And it’s a breakthrough for generalist robots, because a robot that can leverage one of its skills to improve it another makes it much more versatile than one that needs separate training for every single task. In fact, 1x world model’s transfer ability even hints at a future where robots might learn holistically. And 1x world model’s predictions are already matching reality remarkably well, with 1x finding a high correlation between the model’s simulated outcomes and its real world results.

For example, if two policies have a 15% success rate gap in the real world, 1x world model with a 70% prediction accuracy can identify an even better policy 90% of the time. And this reliability also lets 1x use the model to evaluate policies on identical starting states for best comparison. For instance, in one experiment, 1x compared two policies using different image encoders, and 1x world model accurately predicted which would perform better with real world tests confirming it. This alignment means 1x can trust the model to guide decisions like choosing the best training checkpoint, or even tweaking its own model architecture.

But 1x world model isn’t infallible, as it currently struggles with objects that it hasn’t seen in training, with the robot failing to predict how it moves when bumped. Additionally, locomotion is yet another weak spot where small errors in leg positioning can accumulate and skew predictions over multiple steps, with these gaps highlighting the model’s resilience on diverse, high quality data when generalizing. And moving forward, 1x world model will work towards handling broader, more ambiguous tasks with more accuracy, but for now, the 1x world model ties together precise control, data-driven learning, multitask transfer, and real world alignment to solve the core robotics challenge of evaluating policies at scale.

Meanwhile, Mid-Journey just introduced its first ever video model, enabling users to convert static images into short animated clips, with the company describing it as an early step towards AI systems capable of real-time 3D world simulation. And to start, the image-to-video feature is accessible via an animate button on Mid-Journey’s website, allowing users to animate any Mid-Journey generated images. Plus, users can even select an automatic mode where the AI determines motion, or else a manual mode where the motion is described through text prompts. Altogether, two motion settings are available, with the first being slow motion, suiting scenes with minimal camera and subject movement, while the second setting is high motion, supporting more dynamic animations, though results may be less accurate.

And as for length, videos start at 5 seconds and can be extended by 4 seconds up to 4 times, with users being able to modify the original image prompt during extensions, and when it comes to images from outside of Mid-Journey, they can also be animated by uploading them as a start frame and then specifying motion via text. But Mid-Journey still hasn’t mentioned resolution, frame rate, or bitrate, plus there’s no built-in upscaling yet, but so far, videos appear to be 480p MP4 files running at 24 frames per second when downloaded. And as for cost, generating a video is approximately 8 times more expensive than generating an image, with each job producing 4 second clips, and this equates to about 1 image’s cost per 1 second of video, with Mid-Journey saying this is 25 times less expensive than similar services.

And while this feature is currently only available on the web interface, for pro-tier subscribers and above, a video relax mode also allows video generation without using fast processing minutes, potentially reducing costs even more. And for now, Mid-Journey views this video model as a foundational step towards integrating video, 3D elements, and real-time processing into a unified platform, with the company aiming to develop systems for real-time world simulation in the near future. And on top of this, insights from the video model will also improve Mid-Journey’s image generation tools. But that’s just one of the latest generative AI developments, as Adobe also just unleashed its Firefly platform on iOS and Android, empowering creators to ideate, generate, and edit images and videos from anywhere.

Additionally, the Firefly mobile app syncs seamlessly with Adobe’s Creative Cloud, enabling smooth workflows from phone to desktop apps like Photoshop and Premiere Pro. Plus, creators can now harness AI models from Adobe, OpenAI, Google, ideogram, LumaAI, Pika, and Runway to craft assets with unmatched flexibility, from lifelike images to 1080p videos and editable vectors, all via simple text prompts. And on top of this, Firefly Boards, which is now in public beta, is also a game changer for team collaboration thanks to an AI-driven moodboarding service, integrating video remixing and iterative edits, and having over 24 billion assets generated globally, Firefly is gaining momentum with content credentials, ensuring transparency to safeguard creator rights.

But if you want to know more about the latest in AI news, make sure to like and subscribe, and check out this video here. [tr:trw].