Figure 03 Humanoid Robot Learns 8 New Autonomous AI Skills (AI NEWS)

Flopping this over its shoulder to free up its hands for skill number two. This is dynamic object handling and here it’s going to take this towel off of its arm and reposition it for cleaning which leads into skill number three. Here it’s going to perform bi-manual manipulation. First, it’ll pick up this container with one hand and use the other to start scooping these blocks into the bin and then we see skill number four. This is a whole body strategy for efficiency where it tucks the container under one arm to free up the other hand for picking up these items on the couch.

Then we see skill number five here. Dynamic object throws right after he puts down this container. You’re gonna see him first pick up this first pillow and then do something very impressive which is to very naturally throw the second pillow into place. This is with no new algorithms just data. Now on to skill number six. This is in-hand reorientation of the remote as well as choosing the correct button to press to turn off that television set and next he’s going into skill number seven reorganizing tools during motion. He’s going to stow this towel under his arm to free up his other hand and then he’ll demo skill number eight which is just foot placement side stepping through this environment.

But what’s even more interesting here is Helix zero two which acts as a three-part robot brain and if you watch this back everything changes because what you’re seeing here is this three-part system working together to reach general intelligence to some level we can already see starting with its system zero which runs at 1,000 Hertz or 1,000 times per second it calculates balance contact and coordination across the entire body to be able to bend over like this and reach for these objects next it gets into system one this is where it translates perception to full-body joint targets in order to grab this towel from its arm and orient it back onto the table and it thinks at a rate with system one at 200 Hertz or 200 computations per second and then we get into system number two.

This is the reasoning layer. This is for scene interpretation language understanding and orienting around its world to figure out what to pick up next for instance and this goes more slowly and figures out how to do what it needs to do whereas system one figures out what to do and system zero runs on a macro level everything together and so with the integration of all three of these systems We have the Helix zero two robot brain Which is able to ingest all the data from figures Go big project and carry out These real-time trajectories and figure out exactly what to do inside of its environment But the question here is what is this robot going to cost and how quickly can it learn another skill? How many trajectories does this robot need to be able to carry on with another ability? And as for this robots important stats it has a payload capacity of 20 kilograms or about 33 pounds and Its weight is about 61 kilograms or 134 pounds with a runtime of 5 hours All electric with a speed of about 1.2 meters per second So make sure to comment down below how much you’d be willing to pay for these stats But physical intelligence is also tackling robot AGI as the team just deployed their PI 0.5 vision language action model on a real excavator Which is a first for heavy equipment outside of a lab setting and this tackles three major barriers The first is the fact that the amount of real world machine data is scarce The second is that interfacing hardware with industrial equipment is difficult And the third is the challenge of adapting the VLA control from joint encoders to joystick based input And to do this the data was collected using only Monocular cameras and a yolo v8 model that visually tracked joystick movements from the operator and mapped them to a 40 action space Then large tasks like trench digging were broken down into repeatable subtasks to boost training signals from varied unstructured environments And the key insight here is that joystick control works for VLAs despite being non-linear and position ambiguous Which is unlike the joint encoders that every lab robot uses So next the team’s goal is to create a single generalized policy trained across multiple machine types tasks and environments within the next month Meanwhile anthropic just released its new code review tool Which dispatches multiple AI agents per pull requests in order to hunt down bugs in parallel Verify its findings and rank them by severity and in internal tests Anthropic’s pull requests receiving a substantial amount of comments jump from 16 percent to 54 percent With large pull requests being flagged at 84 percent of the time averaging 7.5 issues And less than 1 percent of those findings were marked incorrect So in one case it caught a one-line change that would have broken production authentication And its review process averages about 20 minutes and costs between 15 to 25 per pull request And while it’s already available in beta final approval stays up to the human And finally another AI tool called open claw just went viral in 2026 and was swiftly acquired by open AI But as a result of this businesses lost the open claw independent AI agent platform So in response to this nvidia is preparing to release nemo claw Which is its open source enterprise grade agent platform built on the nemo framework nematron models and nim micro services And it’s expected to be announced at gtc 2026 next week But most importantly nemo tron is hardware agnostic meaning it runs on nvidia amd and infel chips alike With the intention to target regulated industries with agents for data processing content generation and decision making while Differentiating itself from open claws consumer routes with compliance versus architecture and native gpu acceleration Anyways, check out these videos here for more of the latest in AI news Like and subscribe and thanks for watching
[tr:trw].