Boston Dynamics Atlas Killer? 32 DoF M1 Superhuman Robot (AI NEWS)

Specifically, it stands about 128 centimeters tall, or 4 feet 2 inches, and it weighs 28 kilograms, or about 62 pounds. Features 25 degrees of freedom here, and something special about its joints we’ll get into in just a second, but 244 newton meters of torque. And it features this battery with 6 amp hours of power, total running time of four hours. So now let’s break down these 25 degrees of freedom. There are two degrees of freedom in the head for pan and tilt, five degrees in the arm, shoulders, elbows, and wrists, and one degree of freedom in the waist for rotation, then six degrees per leg across the hip, knee, and ankle for walking.

And when it comes to artificial intelligence, the robot is using an onboard Intel Core i7 for compute, and this is to power the stereo RGB sensor array in the head, which enables the robot’s depth perception and 3D environmental mapping. Then, looking down at its legs, there’s the nine-axis IMU sensors, and this provides real-time orientation data onboard with the accelerometer, gyroscope, and magnetometer. But let’s take another look at one of the really unique features here, which is the projection display built into the robot’s face. This allows it to share content for viewing, or educational displays, or even enhanced emotional expressions.

And it also features full color customization, but more important is the robot’s weight at 28 kilograms, or about 62 pounds. This is actually significantly less than Boston Dynamics Atlas, which is 80 kilograms, or 176 pounds. So it looks like they’re positioning themselves in the market to serve as a humanoid companion. But there is another product, the M1, which I’m going to get into in just a second. This is the larger general-purpose humanoid. But when it comes to the pricing specs of the C1, it’s still unknown, but you could expect this would probably be closer to the $5,000 to $10,000 range.

Then, looking at the timeline, when will this release? So far, no dates on the website, possibly 2026 or later, but it looks like this is in an earlier stage of development. And while it doesn’t have hands, the M1 that I’m about to show you does. But speaking of hands, are robots coming to Automate Sports 2? Because the researchers from Fizz H.O.I. or a physics-based imitation and dynamic human-based object interaction may think so. And they’ve just released footage of their unitry G1 humanoid handling all kinds of basketball moves from jump shots to running shots.

But some people are questioning the authenticity of these videos. So is this real? And if so, should robots be able to compete against humans in sports? Or should there be human-only leagues and robot-only leagues that remain completely separate? Anyways, back to the meat and potatoes here, the 5-bot M1. This is a high-performance, general-purpose humanoid. Let’s get into the specs. So it stands at 172 centimeters tall, or about 5 feet 8 inches. And it weighs less than 60 kilograms, around 132 pounds. And it features 32 degrees of freedom in all, with 2 degrees of freedom in the head, 7 per arm, 2 in the waist, and 7 per leg.

But this doesn’t also count for its dexterity. You can see here that it has 5-fingered hands. So these 32 degrees of freedom likely exclude the hands. But looking at the payload capacity, it has a dual arm load of 20 kilograms, which is about 44 pounds, or over 50 kilograms, which is about 110 pounds for backpack capacity, carrying a load on its back. And it can carry up to 83% of its own body weight. And this is partially thanks to its superhuman strength motors, which we’ll get into in just a second. But first, it has a walking speed of over 2 meters per second.

That’s over 6 feet per second, which equals 4.5 miles per hour. And this is similar to a human jogging pace, which is faster than most commercial humanoids currently. But what’s really incredible here is the robot’s peak joint torque of 530 newton meters, which actually surpasses the power of a Formula 1 race car engine. And this is thanks to FiBot’s proprietary FiArc series quasi direct drive integrated cycloid joints. And these basically just offer real time position feedback and error free execution. And they also feature an all in one modular design with the drive motor reducer and dual encoders all on board for easier maintenance.

And the run time for this robot is over two hours continuously with two swappable batteries, and they’re hot swappable within 15 seconds, they can be changed out. And this is to power the onboard Nvidia Jetson Oren NX, which provides 100 tera operations per second of AI performance to the robot for real time neural network processing. And it’s also to power the AI in the 3D LiDAR and stereo RGB cameras, as well as high dynamic range cameras for environmental mapping. And when it comes to control algorithms, the M1 uses self developed motion control software integrated with its proprietary hardware.

And in comparison to the C1, it stands significantly taller with the target application not just for home robotics, but also for industrial logistics and service roles. But there are some information gaps remaining. So far, we don’t know how much will this robot cost? Tell us in the comments down below if you think that this will be a 10 2030,000 dollar robot to compete with the Tesla and bigger robots or if you think it’ll be a much more expensive robot like Boston Dynamics Atlas. Also, there’s no word on when this robot will begin selling.

So probably a 2026 2027 timeline for release. And in terms of software platforms, the operating system and programming interfaces are yet undefined. And finally, flexion reflect v zero might be the coolest modular robot autonomy system so far. This uses natural language to reason through a modular toolkit. And it overcomes the training bottleneck of manually collecting data. And instead, it uses this modular approach with a skill toolbox. Here you can see it knows to use this toolbox to lean down and pick up the toy on its left. And it does this with a three layer architecture at first is with the LLM VLM agent.

And this decomposes tasks from plain English as you see here to select appropriate tools and then maintain spatial context. Then the motor generator creates short horizon, collision aware trajectories from visual input and task instructions. Then finally, the RL based tracker does this, it executes the commands as you can see here across all terrain types, using reinforcement learning train control. And the key advantage here is the clean interface between layers. So they’re each testable independently. And the basic data strategy here is that they’re leveraging simulation data and synthetic data first and then adding real world data only for specific gaps.

So the breadth comes from simulation and the precision comes from targeted real world refinement. Anyways, like and subscribe and tell us in the comments below which tasks this robot should learn next. And thanks for watching. Bye. [tr:trw].