New GEN 3 Humanoid Robot Full Body E-SKIN Does This (AI NEWS)

And it all starts with textile-based electronic skin, a flexible foldable material with a folding radius of under 0.2 millimeters that can be tailored to a robot’s frame like a fitted suit. So, at the fingertips, high-sensitivity tactile sensors pack dozens of sensing elements per square centimeter, detecting gram-level contact and distinguishing surface textures. And that precision now extends across the entire hand. Full-palm integrated tactile sensing captures pressure distribution simultaneously, giving robots a complete map of how an object sits in their grasp. This feeds into dynamic grip force monitoring, where real-time feedback allows continuous adjustment, loosening on a paper cup or firming on a metal tool without pre-programmed force limits.

And underlying all of it is high-precision pressure calibration, ensuring accurate readings across a wide range of forces for robots to be able to handle anything from surgical instruments to industrial parts. But sensitivity doesn’t stop at solid contact. Airflow dynamic sensing detects forces as subtle as 0.01 newtons, fine enough to register a light current of air across the sensor surface. And when it comes to its hands, they’re not just sensing, they’re also communicating. Full-palm post-hell operation translates a human operator’s hand shape and finger positions onto a robotic hand in real time, enabling intuitive remote control, as well as knuckle angle tracking, which captures the bend of each individual joint, providing the motion data needed to replicate the full range of human hand articulation.

And perhaps most notably, this technology has already moved beyond the lab. Manufacturers have begun custom production and mass manufacturing of the electronic skin system, scaling from prototype to commercial supply, with the global market for flexible sensors being projected to reach billions of dollars by the end of this decade. So what’s emerging here is a generation of robots that can perceive the world through sustained physical contact. These are machines that can feel a handshake, catch a slipping glass, or even sense the people around them. But what if you could build a robot that could wear that skin? Menlo Research has just open sourced its Asimov, an agent-native humanoid robot available as a DIY kit now open for pre-orders at a target price of just $15,000.

Named after Isaac Asimov, the science fiction author whose three laws of robotics remain a foundational framework for safe machine behavior, the robot is designed to augment human capability rather than replace it. And Asimov runs on the Menlo OS stack, which integrates five components. Menlo OS provides the agent abstraction layer, the agent platform deploys AI agents with safety guarantees, Uranus validates behavior and simulation before physical deployment, Cyclotron trains motor skills like locomotion and manipulation, and the data engine captures operational data for continuous improvement. And when it comes to the hardware, they’re using off-the-shelf parts and 3D printable components with standardized servo interfaces, an integrated perception stack, onboard compute for low latency reflex loops, and hot swappable battery modules.

Plus, safety constraints and telemetry are implemented at the hardware level where every policy passes through safety boundaries before reaching the robot. So rather than vertically integrating manufacturing, Menlo opened Asimov as a reference design for an ecosystem of manufacturing partners, with Menlo OS abstracting away hardware differences. And here’s an option for the open source robot brain, as researchers just open sourced a framework that lets AI write its own robot control code. So the team from NVIDIA, UC Berkeley, Stanford, and Carnegie Mellon just open sourced CapEx, a framework for building AI agents that control robots by writing and executing code in real time.

Rather than relying solely on pre-trained motion policies, CapEx equips robots with perception APIs, control APIs, and auto-synthesize skill libraries, allowing them to solve zero-shot manipulation tasks that learn policies would generally struggle with. And the framework includes four components. CapGym provides 187 manipulation tasks across tabletop, bimanual, and mobile settings in both simulation and real-world environments. CapBench evaluates 12 frontier models, including Gemini, GPT, Claude Opus, Quinn, and DeepSeek. Next, CapAgent0 is a training-free system that matches or exceeds human expert code on four out of seven tasks without specific tuning. And finally, CapRL applies reinforcement learning directly to the coding agent, a 7 billion parameter model that jumped from 20% to 72% success after just 50 training iterations, programs transferring to real robots with minimal sim to real gap.

Furthermore, this project traces its lineage all the way back to Voyager, an agentic AI built three years ago by the same team to play and learn continuously in Minecraft. So its core ideas, skill libraries, self-reflection loops, and in-context planning, have now graduated from the game world into physical robotics. But while AI is rewriting what’s physically possible, Netflix just taught a model to erase objects from video and rewrite physics they left behind. So Netflix just open-sourced Void, or Video Object and Interaction Deletion, which is a new AI framework that removes objects from video footage and automatically reconstructs how the remaining scene would physically behave without them.

So remove a person holding a guitar, and the guitar falls. Remove a car from a collision, and the debris, smoke, and impact disappear with it. And Void was made possible by building on top of Alibaba’s Kog Video X, which is fine-tuned with synthetic data from Google’s Kubrick and Adobe’s Humoto. Plus, Google’s Gemini Pro analyzes each scene to identify afflicted areas, while Meta’s Sam 2 segments the objects for removal. And finally, researchers just wove artificial muscles directly into fabric, and they accomplished this using electrofluidic fiber muscles. These are artificial muscle fibers that can be woven directly into garments.

They’re 2 millimeters thick, and they use integrated millimeter scale pumps to move fluid through a closed circuit, eliminating bulky compressors. And the result is silent, vibration-free actuation at a power density comparable to the human skeletal muscles. So in demonstrations, bundled fiber lifted 4 kilograms, or 200 times their own weight, while a woven muscle bent a robotic arm 40 degrees, yet remained compliant enough for a handshake. Furthermore, their performance scales by adjusting the ratio of pumps to actuators. And this was all published in Science Robotics in 2026, with the work positioning these fibers as a foundation for consumer wearables, as well as soft robotics that are powerful, portable, and potentially machine washable.

So is this the future of robotic muscles? Comment down below, and don’t forget to like and subscribe, and click on this video here for more on the latest in AI and robotics news. [tr:trw].