New ALLEX AI Robot With 48 DOF Changing Humanoid Dexterity Forever (GEN AI AGENT)

➡ We Robotics and RL World have introduced a new humanoid robot named Alex, which is designed for safe and precise interaction with humans. Alex’s unique features include a next-gen torque system for smooth movement, whole-body sensing, and a highly dexterous hand. A new framework called Beyond Mimic allows Alex to learn human-like movements, and can even handle new tasks without retraining. In addition, Hun Yuan Gamecraft has developed a system for more realistic and efficient control in video games, using advanced techniques to reduce computational demands while maintaining consistency.

Are AI robots finally surpassing human level agility? And what will they do next? Today on AI News, We Robotics, in partnership with RL World, just unveiled its humanoid robot named Alex, but with a special twist to set a new standard in dexterous human-robot interaction. To start, Alex is engineered with a focus on safety, precision and adaptability to be well-suited for environments where collaboration with humans is essential. But where Alex really does things differently is in its next-gen torque system. The robot delivers a peak torque of greater than 120 Nm while maintaining a back-driveable torque of less than 1 Nm.

This balance ensures robust performance alongside smooth and compliant movement, reducing risk when working closely with people. Plus, the mass under the shoulder comes in at under 5kg, contributing to an inherently lightweight and safe profile. But Alex also features whole-body kinesthetic sensing, allowing it to detect and respond to force throughout its entire frame. What’s even more impressive is that the robot’s hand is the first true back-driveable human-like design, featuring 15 degrees of freedom for exceptional dexterity as well as a lifting capacity of up to 30kg, with over 40 N of fingertip force.

Furthermore, its force sensing of less than 100g enables Alex to gently interact with delicate objects. This in part allows for its precise repeatability, achieving results within 0.3mm, and with 48 degrees of freedom across its body, including 7 per arm, 15 per hand, plus neck and waist articulation, Alex is built for responsive natural movements in any collaborative setting. But the next level of humanoid agility just arrived with Beyond Mimic, a breakthrough framework designed to teach robots versatile, naturalistic movement by directly learning from human motions. Unlike prior methods that struggled with translating complex dynamic kinematic references into robust hardware performance, Beyond Mimic introduces a scalable motion tracking system capable of reproducing advanced skills such as spins, sprints and even cartwheels, all while maintaining high motion fidelity.

On top of this, Beyond Mimic debuts a unified diffusion policy that empowers robots with zero-shot, task-specific control at test time. This means the robot can handle new challenges, including waypoint navigation, joystick teleoperation and obstacle avoidance, all by simply following intuitive cost functions with no retraining required. Furthermore, by bridging the gap between simulation and real-world deployment, Beyond Mimic seamlessly transfers newly learned AI skills from virtual environments to physical robots. The results are bold, real-world demos which are promising greater flexibility in whole-body humanoid control by blending motion primitives on demand. And while Beyond Mimic is reshaping embodied AI, to tackle the challenges of long-term scene consistency, efficiency and adaptability across game genres, Hun Yuan Gamecraft just introduced its unified approach to fine-grained player action control, where standard keyboard and mouse inputs are mapped into a shared camera representational space, allowing for smooth interpolation between complex camera movements and in-game actions.

This framework enables not just fluid control, but also more realistic depiction of various gameplay operations. Plus, the model leverages a hybrid history-conditioned training strategy that extends video sequences auto-regressively, ensuring crucial game scene information is consistently preserved, even as players make significant movements or navigate complex scenarios. On top of this, Hun Yuan Gamecraft employs advanced model distillation techniques, reducing computational demands while maintaining visual and temporal consistency over longer sequences. This makes the technology suitable for real-time use in interactive environments, and you won’t believe how it works. The engine is trained on a dataset comprising over one million gameplay recordings across more than 100 AAA titles, providing broad genre coverage and diversity.

Furthermore, fine-tuning on a meticulously annotated synthetic dataset further sharpens precision and action controllability. The resulting video outputs exhibit notable improvements in realism, visual fidelity, and user-directed control. Extensive testing shows that Hun Yuan Gamecraft significantly outperforms previous models in terms of continuity, consistency, and playability. The system supports multi-action visualization, capable of handling complex, sequential player commands, and generalizes to third-person perspectives, ensuring natural controls across diverse game types. With these capabilities, Hun Yuan Gamecraft is setting a new benchmark for interactive video game generation for dynamic and lifelike game experiences. At the same time, Metafare has just introduced Dino V3 as a breakthrough in self-supervised learning for visual understanding.

With this new release, Dino V3 pushes the scale of self-supervised learning for images even further, resulting in what the company calls its strongest universal vision backbones to date. Its core technology is built on extracting rich, dense image features that consistently show strong self-similarities over time. Plus, these dense features demonstrate reliable alignment, not only across different objects, but also under significant style shifts, providing stability and adaptability for downstream tasks. But that’s only the beginning, because Dino V3’s approach is also highly efficient with zero-shot segmentation tracking or few-shot fine-grained segmentation models requiring only minimal manual annotation.

As a result, Dino V3 achieves outstanding performance across a wide array of vision-related tasks, even launching with a specialized vision backbone designed specifically for satellite imagery analysis. Its open-source release includes comprehensive training code, model weights, several efficient model architectures, and tutorials to support new users. And in the latest breakthrough in AI video editing, Runway has released Aleph to push the boundaries of what is possible with multitask visual generation. At its core, Runway Aleph enables creators to edit, transform, and generate video using intuitive prompts, offering a suite of advanced features previously out of reach for most users.

One key highlight is its ability to generate entirely new angles from existing footage, allowing for dynamic storytelling without reshoots. Plus, users can create seamless next shots in a sequence, ensuring narrative continuity with just a simple request. On top of this, Runway Aleph excels at applying any visual style to footage, transforming the aesthetic of a video to match creative goals. Aleph can even alter environments, locations, seasons, and the time of day, providing professional-grade visual effects with minimal effort. And for content creators needing to add or remove objects, Aleph integrates or erases elements naturally, maintaining realistic lighting, shadows, and perspective.

Additionally, Runway Aleph supports complex object swaps, motion transfers from video to image, and the ability to change a character’s appearance, all via straightforward prompts or reference images. Color grading and relighting are also streamlined, enabling adjustments to mood and atmosphere across a scene. Finally, advanced green screen capabilities allow precise background isolation, ensuring flexibility for compositing work. And in AI Agent News, Manus has just announced two new updates, aiming to streamline creative workflows and automate routine monitoring. For creative teams, the company showcased how its AI Agent can now transform written scripts into visual storyboards within minutes.

As a result, campaign development moves from concept to execution at a much faster pace. On top of this, Manus AI introduced the scheduled task feature, designed to simplify staying current with launches, updates, and industry trends. Users can now monitor multiple sources automatically, giving them a comprehensive overview without the daily hassle of manual tracking. Anyways, make sure to like and subscribe for more of the most important AI news, and check out this video here. [tr:trw].