World First AI Robot Gym Trains NEURAs 4NE1 Humanoid To Do This

And the concept here is quite straightforward. When one robot successfully learns a task at the NeuroGym, that skill can be deployed to other robots worldwide instantly. In fact, Neuro combines this real-world gym data with more synthetic training to create what they describe as highly complex transferable models. And companies can already book space in the gym to train applications specific to their needs by either working with Neuro’s trainers or by bringing their own teams while being able to use Neuro’s complete infrastructure. This is why NeuroRobotics has already shifted their focus from operating a single facility to scaling the concept across multiple global locations.

But the company also just announced several upgrades to its for-any-one humanoid. The robot is now designed for both industrial and domestic environments. And it features full-body sensor skin and 360-degree perception. It stands 180 centimeters tall and weighs 80 kilograms, allowing for-any-one to walk up to 5 kilometers per hour while handling payloads between 10 and 100 kilograms. And the robot’s brains run on multimodal AI combining language models, computer vision, and reinforcement learning to navigate unstructured spaces, interpret complex requests, and respond to complex commands in real time. And in another world first, Deep Robotics just unveiled its new DR-02, claiming it’s the world’s first all-weather industrial humanoid.

As for specs, the robot is 175 centimeters tall and weighs 75 kilograms, and now has an IP66 waterproof and dustproof certification. And it’s capable of operating in temperatures between minus 20 to 55 degrees Celsius. And as for flexibility, the robot has 31 degrees of freedom and has a reaching distance of 68 centimeters with each arm. In fact, Deep Robotics says the DR-02 is designed to work continuously throughout rain-soaked outdoor sites, freezing cold storage, and high-temperature workshops. But when it comes to gripping, Boston Dynamics just released new details about its Atlas gripper system, stating its 7-degree-of-freedom hands feature tactile-sensing fingertips, as well as palm-mounted cameras, allowing Atlas to handle everything from heavy automotive parts to delicate objects like coffee cups, and the gripper can even detect when it’s lost its grip or dropped an object entirely.

But when it comes to the three-fingered design, Boston Dynamics explained why they avoided mimicking human hands. They said that while human hands offer exceptional dexterity and fine motor control, it was just too steep of an engineering challenge for now with the heavy payloads that the robot has to handle. So instead, their current three-finger design prioritizes durability, with three fingers with an opposable thumb provide manipulation flexibility while minimizing potential failure points, but Boston Dynamics says their future designs could still evolve. And as for specs, each finger can bend inward and backwards 90 degrees, and because Atlas has no dominant hand, it just uses whichever hand provides the best angle and the most reliable grip based on its position and the object.

But to push robots even further towards generalizability, researchers just introduced a brand new system called FizzHSI, which lets robots interact naturally with the real world. The challenge here is in combining generalizable motion with robust scene perception, which are two capabilities that have previously been developed separately but are rarely combined effectively. And to accomplish this, FizzHSI uses a two-stage process, simulation training followed by real-world deployment. So during the training phase, researchers start with motion capture data of humans performing everyday tasks, like carrying boxes, sitting in chairs, lying down, or standing up.

Next, human motion data is retargeted to human body mechanics with contact points between the robot and objects annotated at keyframes. Then the AI learns to distinguish between robot-generated movements and natural human reference motions. And this adversarial training pushes the policy to produce more lifelike movements while still completing tasks successfully. And for real-world deployment, FizzHSI introduces a course-defined object localization system where operators can manually initialize the approximate object position using LIDAR visualization. So a rough guess of the object’s position is combined with the robot’s movement data to keep tracking things even when they’re out of the camera’s sight.

And finally, once the target pops into view, the Apriltag visual system jumps in for automatic positioning during the interaction. But for even greater dexterity, AGIbot just released its newest OmniHAN, which weighs only 500 grams and has an extremely small form factor of just 180 millimeters or about seven inches. And the OmniHAN features 16 degrees of freedom for ultra-human-like movements, plus it can even sense touch on the back. And this is thanks to the fact that OmniHAN has over 400 sensors built in and they’re paired together with its anti-pinch technology, and all of this allows the OmniHAN to perform pick-and-place tasks with submillimeter-level precision, while the HAN features a universal mounting system so it can be attached to different robots.

But there’s another brand new robot HAN from South Korean company WeRobotics in collaboration with RealWorld as they just demoed their Alex robot autonomously opening a bottle and pouring milk, which is a deceptively complex task for robots. And their partner RealWorld is currently developing foundation models that give robots human-like vision, reasoning, and hand dexterity with a focus on delicate manipulation and dynamic adaptation for real-world autonomy at scale. In fact, when the team studied human hand structure and tool usage, they identified that their bottleneck was insufficient high-quality training data. So to build deployable solutions, they needed to blend human action data with precise scene understanding, object detection, 3D spatial awareness, and multimodal sensor fusion.

And that’s why RealWorld is now turning to robot-generated training data over pure human demos, and their reasoning is that humanoids don’t need a strict human blueprint. But there’s another brand new robot from South Korea that was just unveiled named CAPEX, which was developed jointly by LG Electronics and Korea Institute of Science and Technology, but with a special twist. That’s because CAPEX runs on LG’s XA1 vision language model for AI processing, paired with domestically produced actuators and a multi-finger hand equipped with tactile sensing. On top of this, LG has already moved its research talent into a dedicated product unit, which indicates commercial intent rather than pure research.

And while not much else is known, the project is part of South Korea’s K Humanoid Alliance strategy to compete with established robotics leaders with LG planning a public reveal in November and commercialization targeted within the next four years. And in the latest vibe coding AI tools, Google just expanded the release of its Opal app builder into 15 new countries. And they launched two months ago in the U.S. as a labs experiment, with early users building far more than Google had originally anticipated. So thanks to feedback received by Google’s users, they’ve added no-code debugging to now run step-by-step in a visual editor.

Also, they’ve added real-time errors with pinpointing exact failures to minimize guesswork. And creation times have gone down to just five seconds with infrastructure overhauls, as well as parallel execution that allows for multi-step workflows running simultaneously, slashing delays by up to 50%. Anyways, like and subscribe for more of the latest in AI news and tell me what you think about this unitry robot doing its latest kung fu moves. And tell me if you think you could win a fight against this robot. [tr:trw].