2026 Humanoid Robot Tech NVIDIAs 8-Exaflop AI News

As for applications, the company sees iron as a crucial extension of the transportation ecosystem, as Xpeng believes that humanoids will soon play a transformative role. But for now, the first-generation models have limited intelligence, with the company expecting the robot era to arrive soon, hinting at a 2026 chat-gpt moment for humanoids, and already, iron is their working prototype that has begun assisting on Xpeng’s P7 Plus electric vehicle production lines. Standing 173 cm tall and weighing 70 kg, iron is currently equipped with over 60 joints and 200 degrees of freedom in total.

But most impressive is its 15-degree of freedom hands that allow iron to perform precise tasks in the real world. As for intelligence, iron’s capabilities are powered by Xpeng’s proprietary Turing AI chip, a 40-core processor designed for real-time reasoning and decision-making. Its AI brain is combined with the Xpeng’s Eagle computer vision system to allow the robot to move autonomously, mimic human-like postures, and even engage in basic conversation. For example, the CEO stated that iron can even operate in German while functioning without an internet connection. And during the IAA keynote, iron accompanied the CEO on stage while fielding questions about flying cars and demonstrating its conversational abilities.

Xpeng expects these developments to accelerate quickly as production and intelligence improve, but there’s another robot that’s already started selling and it’s called the GR3 from 65 cm tall and weighs 55 kg. For movement, the GR3 features 55 articulated joints, and to power it all is Fourier’s proprietary full-perception multimodal interaction system. Unlike traditional robots, this system merges visual, audio and tactile feedback into a real-time emotional processing engine. On top of this, the GR3 is outfitted with 31 pressure sensors for nuanced tactile feedback, allowing the robot to detect and respond to touch with lifelike social cues.

For instance, animated micro-expressions, blinking, eye-tracking, and gesture responses are triggered when the robot perceives human contact. This is partially thanks to the GR3’s visual recognition system, which is powered by an RGB and structured light camera module. This setup provides reliable facial recognition and continuous dynamic tracking so that users always remain within the robot’s visual field. Next, the audio system leverages a 4-microphone array for omnidirectional voice activation, echo cancellation, and precise sound source localization. This allows the GR3 to orient itself towards speakers and maintain eye contact for a better conversational flow. And the GR3 supports multiple mobility modes, including straight leg walking, running, a lively walk, and even an energy-saving fatigue mode for long-term operation.

As for power, the robot’s battery runs for three hours of continuous use, and on the design front, the GR3 avoids an industrial gunmetal color scheme with a soft touch exterior, cushioned surfaces, and premium automotive upholstery in warm neutral tones. And in the near future, Fourier anticipates expanding the GR3’s role into assistive care to support mobility, health monitoring, and rehabilitation in both home and clinical settings. This comes as figure’s artificial intelligence just taught itself how to load a dishwasher after having previously learned how to fold laundry and sort packages. But what sets this achievement apart is that the exact same core model was used, with no new algorithms or special engineering needed, just additional data.

And while loading a dishwasher looks simple, it actually involves a series of complex robotics problems. First, helix must pick up dishes from cluttered stacks, reorient or hand off fragile objects using two arms, and place items precisely in racks that leave only centimeter scale margins for error. Next, every dishwasher load has unique objects, unexpected arrangements, and possible collisions, requiring the system to constantly adjust and recover to succeed. Impressively, helix learn to singulate stacked plates, reorient glasses between hands, and even adapt to messy situations. But is it ready for prime time to compete with your girlfriend in a face-off of the dishes? Tell us in the comments below who you think would do a better job.

Meanwhile, Tesla just revealed a gold-colored version of its Optimus humanoid robot. However, Elon Musk was quick to clarify that it’s just an intermediate upgrade for version 2.5 rather than a full third-generation release as some had speculated. As for the optimizations, they’re primarily external. The Optimus 2.5 now features smoother shell edges, better covered joints, and significantly reduced seam visibility compared to version 2. Plus, the video demonstration offered only incremental glimpses of progress, leaving out any evidence of advanced autonomy or dexterous object manipulation so far, ending the video before the Optimus attempted to retrieve anything.

Tesla emphasized that the update’s improvements are about refinement and polish rather than introducing new functions or capabilities as the company planned a commercial release for some time in 2026. For now, though, it’s unclear if they’re on schedule or if the level of AI autonomy has reached the bar they were aiming for, and in the newest AI hardware announcement from NVIDIA, the Ruben CPX has been unveiled. This new graphics processing unit is specifically built for handling enormous context windows, paving the way for artificial intelligence systems to efficiently manage million token coding tasks and generative video workloads with unmatched speed and efficiency.

The Ruben CPX works seamlessly alongside NVIDIA Vira central processing units and Ruben graphics processing units within the newly launched NVIDIA Vira Ruben NVL 144 CPX platform. This integrated NVIDIA MGX system delivers a remarkable eight exaflops of artificial intelligence compute power, offering seven and a half times more performance than NVIDIA’s previous systems. On top of this, the platform includes 100 terabytes of high-speed memory and achieves 1.7 petabytes per second of memory bandwidth, all within a single rack. And for an even more efficient design, Ruben CPX integrates advanced video decoders, encoders, and long-context inference in a single chip.

Furthermore, the Ruben CPX utilizes a cost-efficient monolithic die design featuring robust NVFP4 computing resources to deliver up to 30 petaflops of compute power with NVFP4 precision and includes 128 gigabytes of GDDR7 memory for demanding workloads. Last but not least, Ruben CPX is fully supported by the NVIDIA AI stack, including the Dynamo platform for efficient inference scaling and the full suite of Nemetron multimodal models. Plus, its attention mechanisms are three times faster than previous NVIDIA platforms. Anyways, like and subscribe for more AI news, and check out this video here. [tr:trw].