Clone Humanoid Robot With 206 BONES + Superhuman Hand (AI NEWS)

➡ Linkerbot has developed a robotic hand, the L30 Phantom, that mimics human hand movements with high precision and speed, useful in tasks requiring delicate handling and precision. Meanwhile, Clone Robotics is creating androids with artificial muscles and bones, aiming to replicate human anatomy. Generalist, an AI company, has introduced Gen 1, a ‘brain’ that can learn and perform a wide range of tasks, improving the adaptability of robots. Lastly, RobotKit is a software platform that upgrades existing robots with autonomous capabilities, including perception, navigation, and reasoning.

A robotic hand just moved its fingers at 450 degrees per second with superhuman precision, but how much can it feel while doing it? Linkerbot, a robotics company specializing in dexterous manipulation, recently introduced its L30 Phantom, a tendon-driven robotic hand that the company says is engineered to mirror the complex biomechanics of human fingers, thumbs, and palms. And the L30 features 22 degrees of freedom, which is a figure that places it among the highest degree of freedom robotic hands currently available. So, its tendon-driven architecture routes cables through the hand’s frame in a manner structurally analogous to how biological tendons transmit force across joints, allowing each finger to flex, extend, and rotate with what the company describes as human-like range of motion.

But on the precision side, the hand achieves repeat positioning accuracy of plus or minus 0.20 millimeters, with core joint speeds reaching up to 450 degrees per second, and gripping forces across the four fingers range from 10 to 20 newtons, while the thumb delivers between 9 to 20 newtons, with a force resolution of 0.5 newtons. And as for its form factor, the entire mechanism fits between a 179 by 85 millimeter space, operates on DC 24 volt power, and communicates through an RS485 robot serial communication protocol or robot operating system control interface. Plus, LinkerBot built the L30 for use cases that extend well beyond conventional pick-and-place robotics.

Companies cites advanced laboratory automation, like tasks including twisting microcaps, operating micropipettes, and handling fragile glassware as their primary application. Plus, other target domains include embodied AI research, where the hand could serve as a physical training platform for reinforcement learning algorithms that are focused on tactile manipulation, as well as assisted medical procedures requiring submillimeter precision and electronics microassembly in confined spaces where traditional robotic arms lack sufficient dexterity. But while the L30 mimics how human hands move, another company is replicating human anatomy itself. Clone Robotics is taking a fundamentally different approach to humanoid robotics.

Rather than assembling servos, gears, and rigid actuators, the company is constructing androids from the inside out, with 206 polymer bones, artificial ligaments, and a proprietary synthetic muscle fiber called myofiber that attaches to anatomically accurate points on the skeleton just like biological muscle does. In fact, myofiber, which Clone says it pioneered in 2021, is produced as monolithic musculotendent units to eliminate failure points. And the company states that each three-gram fiber responds in under 50 milliseconds, delivers over 30% unloaded contraction, and generates at least one kilogram of contraction force. And Clone describes this as the only artificial muscle currently capable of achieving that combination of weight, power density, speed, and energy efficiency.

Plus, the skeletal system mirrors the full human anatomy with fully articulated joints held together by artificial ligaments and connective tissues. And the shoulder complex alone has 20 degrees of freedom across four joints, plus with each hand, wrist, and elbow contributing 26. And in total, just the upper torso without legs possesses 164 degrees of freedom. And to power the muscular system is what Clone calls the most sophisticated hydraulic system ever designed for a humanoid. A 500 watt electric pump comparable to the size of the human heart delivers liquid at 40 standard liters per minute and 100 psi to supply the entire body.

Then, Clone’s miniaturized aqua jet valves measure just 12 millimeters and combine 100 psi water pressure with 2.28 standard liters per minute flow at under one watt of power consumption. And the nervous system runs on 70 inertial sensors for joint level proprioception, 320 pressure sensors for muscle level force feedback, and four depth cameras mounted in the skull. Sensor data is fused along control boards mounted on the vertebrae and processed by an NVIDIA Jetson Thor GPU running CyberNet, which is Clone’s visual motor foundation model. And the company’s latest iteration, called Torso 3, is its very first hydraulic Torso, which features hundreds of in-house valves that triple the hydraulic power output at one third of the power consumption and 10 times lower cost than the previous off-the-shelf components.

So, Clone is building the body, but a body still needs a brain, and one company thinks it just built the first one that can actually use it. So, Generalist, a robotics AI company, recently unveiled their Gen 1, which they describe as a generalist foundation model trained from scratch on half a million hours of physical experience. And the company says that Gen 1 is designed to function as a single brain for many different robots, enabling them to understand and act in the physical world across a broad range of tasks. And according to the company, Gen 1 can already fold laundry, perform industrial kitting of automotive parts, and even service other robots.

But Generalist emphasizes that no single task is the point. Rather, it’s the ability to learn many tasks well and acquire new ones in a short amount of time. To this, the company reports that Gen 1 can also learn from experience to complete tasks up to three times faster than what it calls the current state of the art. And with Gen 1, Generalist says they’ve already scaled further and introduced innovations across the stack. So, Generalist frames this mastery as the combination of three capabilities, reliability, speed, and improvisation. And while reliability and speed have already existed in industrial robotics since the 1960s, Generalist argues that improvisation, open-ended problem-solving learned through physical interaction, is what has been missing this whole time.

So, Generalist compares Gen 1’s emergent problem-solving abilities to the experience of using a large language model for the first time and seeing it connect ideas in unexpected ways. This is why Generalist views Gen 1 as capturing a dimension of intelligence that’s absent from current tech space AI, the adaptive, hands-on reasoning that comes from acting in a messy, unpredictable world. But not every robot ships with a brain like Gen 1. So, what happens when you need to give an off-the-shelf robot intelligence that it wasn’t built with? That’s the problem RobotKit is designed to solve, as rather than building a foundation model or a humanoid from scratch, RobotKit is a software platform that turns existing robots into autonomous agents by layering perception, navigation, manipulation, and agentic reasoning on top of the hardware that already exists.

So, the platform is built on the ROS2 standard, or robot operating system, meaning that it supports any robot in that ecosystem. And the core version enables a robot to understand the space it occupies, autonomously navigate within it, handle basic pick-and-place manipulation tasks, and reason independently through agentic workflows. And on the perception side, RobotKit combines YOLO-based object detection for real-time localization, vision language models for spatial reasoning about how objects relate to each other in the environment, and temporal relationship processing to understand how tasks and objects change over time. Then, navigation is handled through autonomous route planning with global and local planners, waypoint guided movement with obstacle avoidance, and frontier exploration that builds spatial memory as the robot moves.

And for even more precision, RobotKit offers cloud-based training through vision language action models, allowing users to teach robots specific skills, executable via natural language commands. Anyways, like and subscribe for more of the latest in AI and robotics news, and thanks for watching. [tr:trw].