Figure 03 Demos Helix 02 Humanoid Robot Automation Gamechanger (AI NEWS)

And just look at the way that it carries itself from one point in the kitchen to another. This is no-teller operation, no remote control, just fully AI-powered. See it open up the pantry there? So far, the figure team has achieved 61 local manipulation actions with zero human intervention and no resets. But the real breakthrough here, as you watch this robot, is in its Helix 2 AI brain that does things that the Helix 1 just couldn’t. So first, last year’s Helix 1 was a single neural network that controlled only the upper body from pixels.

But now, as you can see here, this robot uses a single neural network to control the entire robot body, just using raw pixels, meaning it sees and interprets the world. As you can see, it’s interpreting what it’s picking up, and Helix 2 actually extends its control to the entire body of the robot for tasks like walking, manipulating, and balancing as part of one continuous system. And because of this, it’s figure’s most capable humanoid robot model yet, and arguably the most capable in the world until now when it comes to autonomous, real-world, useful tasks.

And something to remember as we watch this robot is that it works on the basis of all sensors in, all actuators out, meaning that every onboard sensor, like for vision to pick up these glasses, touch to feel them, and proprioception to figure out where it is in space, they all connect directly to every actuator. And they all go through a single unified visual motor neural network. So that means there are no handoffs between separate systems, which allow for this very smooth natural motion, rather than the stereotypical jerky robot movement that we were used to.

And the real game changer is in figure’s new approach with its three system architecture instead of two. So let me tell you what I mean. So first there’s system zero, which is the foundation layer that allows this robot to basically use a foundation model for human-like whole body control. This lets the robot know how humans move, which allows it to move so smoothly as we see in this demo, maintaining its balance and stability regardless of its stance. You can see here it’s just leaning down to put in that soap. And it’s actually the backbone of physical embodiment.

Look at the way that it uses its foot to just pick up that door and close the dishwasher. And because of this, it ensures that every motion is smooth, safe, and stable. In fact, system zero was trained over 1,000 hours of joint level retargeted human motion data. And rather than engineering separate reward functions for walking, turning, crouching, or reaching, it just learns to track human motion directly. And the kicker here is that it’s trained entirely in simulation across 200,000 parallel environments. Plus, thanks to extensive domain randomization, it even allows for the direct transfer of skills to real robots as well as generalization across the fleet.

And in terms of raw tech specs, it’s working on a 10 million parameter neural network, which takes full body joint state and base motion as input. And then it outputs joint level actuator commands at 1000 Hertz, or about 1000 times per second. And the results of this can be seen pretty clearly in these different movements of the robot crouching, leaning over, doing these lunges, you can really tell the fluidity of the system. But what really gets impressive is when you see these dexterous tasks being carried out. For instance, we look at tactile sensing, and the use of the palm cameras for these manipulation tasks that are just beyond anything that we’ve seen so far.

So first, they demonstrate the unscrewing of a bottle cap. And something difficult for the robot is that it has to first stabilize the bottle while applying continuous control rotation to be able to remove this cap without losing its grip or crushing the container. And this requires by manual coordination, meaning both hands have to maintain a sturdy grip force, as well as steady torque control. Next, let’s take a look at this ultra dexterous task of pulling out a pill from a medicine box. So the robot has to locate and extract this single small pill from the organizer.

And the pill is actually occluded, meaning it’s hidden from the camera view from the head. And so this requires palm level visual feedback and tactile guided precision grasping. Then we see the robot using a syringe. And this is actually quite impressive because it has to dispense a very precise volume, all while there are variable resistance and tolerances. So this requires force controlled actuation with tactile feedback and coordinated multi finger stabilization. So the robot has to figure it out as it goes. And finally, we have this example of picking and placing metal pieces.

This is from the BotQ manufacturing facility. And here, the robot has to pick up these objects that are partially covered by each other and shifted during their interaction and transfer them from one container to another. So nothing too crazy there. But there are a few things that you probably missed throughout this demonstration. So we’re going to take another look back, starting with the robot closing the drawer with its hips when its hands were occupied. Next, the robot was able to lift the dishwasher door with its foot, meaning that the robot is able to use its entire body as a tool, not just its hands.

Then we look at bimanual coordination. So in terms of objects being picked up and transferred between hands stacked and placed, both arms are operating as a coordinated system throughout. And in terms of the scale of motor control, the same neural network is producing millimeter scale finger motions and room scale walking motions. And all of this is dynamically ranged spanning four orders of magnitude. And this is a big deal because in terms of local manipulation, this has been robotics hardest unsolved problem for decades, not just because walking or manipulation is hard alone, but because doing both of them together resists clean decomposition.

So for instance, when the robot lifts something, its balance changes, or when it steps forward, its reach changes. But make sure to tell us down below whether or not you would want a fully built closed source proprietary humanoid robot or an open source robot that you could control, but it could be a little bit more elbow grease to get it up and running. If so, you’ll probably like this video here. Anyways, like and subscribe for more AI news and thanks for watching. Bye. [tr:trw].