Mirage Generative AI Gaming Is CRAZY New 54 DoF Humanoid Robot

➡ Dynamics Labs has launched Mirage, a revolutionary AI-powered gaming platform that allows players to create and interact with dynamic 3D worlds in real time. Unlike traditional games, Mirage offers unique experiences for each playthrough, with the ability to mold the environment using simple commands. The platform, developed by experts from top tech companies and universities, uses advanced AI techniques to generate realistic, flexible content based on user prompts. In other AI news, Roque Robotics and Agility Robotics have showcased impressive advancements in humanoid robots, while AI researchers have introduced VITACformer, a system that enhances robots’ ability to perform complex tasks by integrating visual and tactile inputs.

Today on AI News, Generative AI Gaming just became accessible to everyone, plus a new humanoid robot just dropped. But first, is Gen AI the future of open-ended PC and console gaming? Dynamics Labs just put their money where their mouth is with the release of Mirage, a real-time generative engine which they say is the world’s first AI native user-generated content game platform that’s powered by live AI world models. And among Mirage’s capabilities are real-time playability at 16 frames per second, dynamic user-generated molding of the environment through natural language commands at any point, extended consistent gameplay sequences of several minutes or more, cloud-based streaming for instant access from any device without downloads, infinite replayability since each world and playthrough is uniquely generated, and multimodal control supporting text, keyboard, and controller inputs.

But let’s start with real-time playability, because unlike traditional game engines, with Mirage anyone can now shape, generate, and interact with fully dynamic worlds as they play, and herein lies Mirage’s first innovation, its real-time world generation. This means that entire 3D environments materialize and evolve on the fly based on user inputs, and Mirage has already released two playable demos. The first demo is called Urban Chaos, and it mirrors the open-ended city exploration of games like Grand Theft Auto, and the second demo is called Coastal Drift, and it’s a free-roaming sandbox game like Forza Horizon.

Both these demos are procedurally generated in real time, meaning that no two sessions are alike, and the experiences are shaped directly by each player’s imagination and choices. Furthermore, this approach is a breakthrough in comparison to recently publicized AI-driven gaming research, like Google’s AI Doom and Genie, or Decart’s AI Minecraft, or Microsoft’s AI Quake 2. Importantly, while those projects showcase AI’s ability to generate rudimentary environments or extend classic games, Mirage instead distinguishes itself in several other ways. First, it enables unrestricted user-generated content, and this lets players alter their world at any moment by typing a request or command, then the game renders their ideas.

Second, Mirage is able to deliver much more realistic visuals than the pixelated or blocky graphics that are typical of current AI-generated titles. On top of this, Mirage supports extended, interactive play sessions, each lasting well over 10 minutes, without losing coherence or visual fidelity. In fact, the Mirage team calls this shift in gaming user-generated content 2.0, because in legacy games, content is static and finite, meaning the city is mapped, the objectives are set, and the experience eventually comes to a close. But with Mirage, everything is open-ended, placing creative agency directly into the hands of the player moment by moment.

And to power Mirage, Dynamics Lab is using an AI architecture that’s centered on large, transformer-based auto-regressive diffusion models. And this technique starts with a vast dataset of diverse gaming scenarios sourced from across the internet, and then they couple it with their own high-fidelity human-recorded gameplay sessions. Next, the data pool is processed via a vertical training pipeline specifically tuned for the gaming domain, which allows the AI to deeply internalize the logic, rules, and interactive patterns of various genres. As a result of this, Mirage is able to generate coherent, realistic, and flexible content in response to user prompts with frame-level prompt processing.

Next, the visual updates are streamed back to the player’s device with a full-duplex communication pipeline that keeps input and output flowing simultaneously to minimize latency. And Mirage is built on a customized transformer that has specialized visual encoders and revised positional encoding, plus a structure for lengthy interactive play sequences. And all of this combined allows Mirage to think and respond intelligently during long, complex gameplay while balancing between generation speed and content quality. And here’s the twist, because in order to maintain the illusion of a persistent, evolving world, Mirage actually relies on an extended context window with key-value caching.

And this just means that the AI remembers the visual and narrative elements throughout the entire play session, ensuring that changes made early on remained consistent and logical in subsequent scenes. But most impressive might be the team behind Mirage, as it’s composed of researchers from Google, Nvidia, Amazon, Sega, Apple, Microsoft, Carnegie Mellon University, and the University of California. And as for the future, the Mirage team believes generative gaming is an entirely new creative medium. For platforms that thrive on user creativity, Mirage could open up immediate personalized content loops, and for broader entertainment ecosystems like YouTube, it delivers formats that can be endlessly replayed, and for online communities, it lowers the barrier to game creation to a single idea or prompt, allowing anyone’s imagination to become a playable reality now.

But beyond generative AI, Roque Robotics just released an embodied intelligence update for its new humanoid robot, called the HumanX, and it features multimodal perception, fusing together pose, vision, force, and tactile sensors, and it has a single-arm payload of 3 kilograms with a single-hand grip torque of 15 newton meters, with a top speed of 0.7 meters per second, and the robot features 54 degrees of freedom in total, but make sure to tell us in the comments how much you would pay for a robot like this. Meanwhile, Agility Robotics put their humanoid robot Digit to the test in a dramatic pull of the rug challenge.

With its vision systems disabled, Digit relied on AI train balance and reflexes to recover from sudden, unexpected slips. Trained through thousands of simulations, Digit reacted faster than a human and stayed upright, showing off animal-like movements and superhuman recovery, and it’s a powerful display of how AI and robotics are merging to handle real-world challenges in work environments, such as robots walking on slippery or uncertain terrains. In Embodied Intelligence, AI researchers just introduced VITACformer, a unified visual-tactile framework designed to enhance dexterous manipulation. Its key innovation is the integration of visual and tactile inputs to enable robots to perform complex tasks like never before, and at the heart of VITACformer is a cross-modal representation that fuses both visual and tactile signals using cross-attention layers within its policy network.

This approach allows the system to create a rich latent space that captures critical interaction dynamics, but even more important for VITACformer is its tactile prediction head, which auto-regressively forecasts future tactile feedback, and unlike traditional methods that rely solely on current touch signals, this predictive capability enhances action relevance, enabling more informed decision-making for manipulation tasks, and VITACformer was tested on four short horizon tasks where it demonstrated superior performance compared to baseline methods like action chunking with transformers. Meanwhile, qualitative comparisons revealed VITACformer’s high task success rate, with the model consistently achieving successful outcomes in scenarios.

But the framework’s most exciting potential is in a long horizon hamburger-making sequence with 11 stages. VITACformer completed all these stages with continuous, high-precision control using an anthropomorphic hand sustaining performance for approximately 2.5 minutes, and this proves the model’s ability to handle intricate, multi-step tasks that demand dexterity and adaptability. By moving beyond passive perception and leveraging predictive tactile feedback, VITACformer is potentially enabling the next line of personal robots that consumers can use to do household tasks like cooking and cleaning, but if this were to work, how much would people be willing to pay for it?
[tr:trw].