Figure Just Built the Smartest AI Robot Ever and Experts Are Shocked

A robot that instantly understands you, sees what you see, and handles objects it's never touched before. No training, no coding, just pure intelligence. Figure's new Helix AI is making that happen. And it might just be the biggest leap in home robotics yet. A few weeks ago, Figure ended its collaboration with OpenAI. Brett Adcock, the founder of Figure look, literally went on Twitter on February 4th to say that the partnership was over because his company had made a huge in house breakthrough in fully end to end AI for robots. That in house breakthrough turned out to be Helix. Right after that announcement, we also found out that Figure was discussing a massive new funding round. Reports say $1.5 billion at a $39.5 billion valuation. That alone signals that investors are pumped, probably because they got a sneak peek at the tech. And honestly, who wouldn't be excited about robots that can put away your groceries? So let's talk about Helix and why it matters. The name itself, Vision Language Action, already tells you it integrates three big components. It sees the environment, understands instructions, language, and then acts on them. Action Figure calls it a generalist model because it's not just for one narrow task. Supposedly it can handle thousands of different household items without needing a ton of demonstration data or specialized coding for each new object. In one of Figure's demo videos, a person basically said, hey, can you two robots put away this bag of groceries? The robots scanned the kitchen, opened the fridge, placed items inside, and even put an apple in a fruit bowl. What's even crazier is that these tasks involved multiple robots cooperating, handing items back and forth with zero fuss. Figures founder Brett Adcock says they've been working on Helix for over a year, aiming to solve general robotics. And by like a human, he means Helix not only can reason through problems, but also respond to voice prompts and handle new objects it's never seen. We're talking about serious object generalization here, which is huge because the average household doesn't look like a tidy assembly line. Households are messy, things change, lighting conditions vary, and every home is basically unique. So to make a robot truly at home in that environment, it needs a level of adaptability that's historically been missing in robotics. One of the biggest bragging points for Helix is that it runs on what they call system 2s2 and system 1.s1s2 is the slower Internet pre trained vision language model part. It's the piece that can interpret natural language commands, look at the scene, and come up with the general plan. S1, on the other hand, is the fast visuomotor policy that translates that plan into real time physical actions at 200 Hz. Think of S2 as the brain that takes a step back to reason about what needs to be done, while S1 is the nervous system doing quick on the fly motor control. The two systems run on low power GPUs on the robot itself itself, making Helix commercially viable right out of the gate without a big external computer. According to Figure, this approach solves a classic trade off. Previously you had large language models that were good at conversation but too slow for moment by moment control. And you had smaller control policies that were super fast but very narrow in scope. By splitting Helix into S2 for slow thinking and S1 for fast acting, they they get the best of both worlds. They even built a DataSet of around 500 hours of teleoperated behaviors. Then an auto labeling VLM was used to generate natural language instructions for training. So each demonstration correlates to how a human might talk. But that's not all. Helix introduces what Figure calls a series of firsts in robotics. The first is full upper body control. We're talking 35 degrees of free freedom. Wrists, torso, head, fingers, running at 200 hertz. That's a lot of joints to coordinate, especially if the robot is picking up objects of different shapes and weights. Or opening a fridge door while twisting its torso to see inside. The second big first is multi robot collaboration. They showed two separate humanoid robots passing grocery items to each other with just verbal commands like pass the bag of cookies to the robot on your right. Both robots interpret the environment and coordinate that handoff. Then there's the pick up anything emergence. Because Helix relies on large language understanding plus integrated visual processing, it apparently doesn't need separate training for every single object. The system can handle thousands of new household items purely through voice instructions. If you say pick up that weird plastic dinosaur, the system can figure it out even if it's never encountered that object before. Another big highlight is how Helix uses a unified neural network approach. Other than having multiple separate modules for different tasks, Helix has a single 7 billion PhD vision language model, the S2 part, and an 80 million PhD transformer for control. This single set of weights is capable of controlling the entire upper body for tasks like picking, placing, opening drawers or handing items to another robot. It's basically a universal approach to a ton of different actions. No fine tuning needed for each new chore in your kitchen. Finally, Helix is said to be commercial ready. It runs on embedded GPUs, so you don't need to be tethered to A giant server in the cloud. That's a pretty big deal if you want a robot that can move around the house without worrying about a monstrous cable or a chunk of hardware in the corner. It also means the system is faster and more secure, less chance of random latency issues if your Internet goes down. Of course, this is all still stage stuff. If you watch the videos that figure put out, you'll see short demos where everything looks smooth, but it's probably the result of tons of behind the scenes testing. It's safe to assume we're a few years away from these robots cooking full meals, folding all your laundry, or vacuuming your entire house the way a human might. But the company is positioning Helix as a game changer that can scale from industrial settings, where they've been working with automotive companies like BMW, to actual home environments now. Interestingly, Meta was reported to be building a new division for AI powered humanoid robots, and another startup, Apptronic, recently raised $350 million with backing from Google DeepMind. And there are even rumors that Apple might jump into the race. So it's basically turning into a robotics arms race among tech giants, each hoping to define what the future of humanoid robots will look like. Brian heder over at TechCrunch pointed out that many humanoid robot companies focus first on industrial and logistics scenarios like moving boxes in a warehouse because the environment is more structured. But Figure is doubling down on the home environment, which is notoriously difficult for robots to navigate. People move furniture small, spill liquids, leave half folded laundry lying around, and do all sorts of random stuff. So if you solve for that, you automatically solve a huge chunk of general use robotics challenges. Another interesting note is that while Helix is a major pivot toward home use, FIGURE hasn't abandoned industrial uses. It's more like they see the home environment as this ultimate test bed for adaptability. If you can handle all the weirdness of a person's house, you can probably handle most industrial tasks as well. And bridging that gap between factory floors and living rooms is something that's been talked about for a long time but rarely achieved because the technical challenges are so different. Helix's approach kind of bypasses this older method of either meticulously coding each action or training a robot on thousands of near identical demos. By using a Vision language model, Helix learns a lot of everyday common sense, like how a fridge door opens or where apples usually go in a kitchen. Then S1 does the fine grained stuff so the robot doesn't need hours of retraining every time it encounters a new mug or toy. This is precisely the kind of approach that could bring a step change in capabilities, to borrow Adcock's words. Now, obviously there are questions about safety, reliability and even hacking. You might be asking, wait, can somebody hijack my home robot and tell it to do weird stuff? That's something we'll probably see more about as the technology moves forward. For now, Figure is focused on the core AI, making sure Helix can generalize fast and handle tasks on the fly. We're also seeing a big push on user friendly features, voice prompts in natural language, a single system that can handle objects it's never encountered, and the capacity for two or more robots to collaborate in real time. According to Figure, Helix is like a human in that it understands speech, reasons through problems, and grasps objects on demand. Not everyone's ready to invite a six digit price humanoid into their home, but as mass production ramps up and the cost hopefully comes down, you might be seeing these robots in more places, perhaps in service sectors, elder care, or even just as fancy household companions. So let me know your thoughts in the comments. Would you trust one of these humanoids to rummage through your kitchen? Or do you think we need to wait until the tech's even more robust and hopefully cheaper before letting a robot grab a glass out of your cupboard? Thanks for watching and I'll see you in the next one.