Figure AI’s NEW Robot Cleans Entire Living Room Autonomously

Figure just dropped something that has the robotics world absolutely speechless. Their Figure-03 robot just walked into a messy living room, and cleaned the entire thing by itself. No human behind a controller, no pre-written scripts, just pure AI. And the crazy part? It picks up a spray bottle, wipes down a table, then whips the towel over its shoulder like it's been doing this its whole life. It tosses a pillow back onto the couch with a casual flick. And then, this is what broke the internet. It picks up the TV remote, rotates it in its own fingers to find the right button, and turns off the TV. All of this powered by Helix-02, Figure's unified AI brain that controls the entire body from head to fingertips. Under 3 minutes. Let's watch the full thing. But let's talk about what's actually going on under the hood, because the real story here isn't the cleaning. It's the AI making all of that possible. This system is called Helix-02, and to understand why it matters, you need to know how robots have traditionally worked. For decades, robotics engineers have taken a modular approach. One system handles walking. Another handles arm movement. A separate controller manages balance. And then you need something to coordinate all of them. It's like having four different drivers fighting over the same steering wheel. If any one of them is slightly off, the whole thing breaks down. What Figure did with Helix-02 is radical. They got rid of all of that. They took over 100,000 lines of manually written control code, and replaced it with a single neural network. One unified AI system that controls the legs, the torso, the head, the arms, and all 10 fingers. Simultaneously. That's why the movement looks so natural. The robot isn't switching between walk mode and grab mode. It's doing everything as one continuous flow. Just like we do. Now the original Helix, version 1, could already do some of this, but only for the upper body. It could react to what it saw and manipulate objects. Helix-02 extends that control to the entire body. Every joint. Every limb. All driven by one system. So how does one neural network manage all of that? It's actually structured as three connected systems working together. Think of them as three layers of the brain. The first layer is like the robot's reflexes. It handles posture, balance, and low-level motor adjustments. Figure trained this layer using over 1,000 hours of recorded human movement, plus reinforcement learning across hundreds of thousands of simulated environments running in parallel. This layer operates at 1,000 cycles per second. So 1,000 times every second, it's checking balance and making tiny corrections. That's why the robot doesn't topple over when it leans forward to wipe a table or shifts its weight to toss a pillow. The second layer connects what the robot sees to what it does. It takes all the sensor data, from the head cameras, the palm cameras, the touch sensors in the fingertips, and translates that into precise joint movements. This runs at 200 cycles per second. While the first layer keeps the robot upright, this layer is figuring out the exact angle and force needed to grab a spray bottle or rotate a remote control in its hand. The third layer is the high-level planner. This is the part that actually understands goals. You can tell it, clean up the living room, and it breaks that down into sub-tasks. Pick up the towel, spray the surface, wipe it down, collect the toys, toss the pillow back. It doesn't need to know the specific joint angles. It just sets the objective and lets the lower layers handle execution. Here's what most people miss about this demo. The impressive part isn't any single action. Picking up a towel? Robots can do that. Spraying a bottle? Sure. The hard part is doing all of it in sequence, for minutes at a time, without failing. In robotics, this is called long-horizon control. A short task has a few steps and a high chance of success, but a long task is dozens of decisions chained together, and each one has to go right. Even a small error early on can snowball and wreck the entire sequence. The robot doesn't get to pause and reset. It has to keep tracking what it's done, adjust on the fly, and recover from minor mistakes. All in real time. This living room demo is exactly that kind of challenge. The robot navigates tight spaces between the coffee table and couch. It switches between tools. It manages objects with both hands. And it does this across a continuous run with no humans stepping in to fix anything. That's the gap between a robot that can perform a trick, and one that can actually finish a job. One thing that really unlocks Helix-02 is the hardware on Figure-03. Specifically, the sensors. Most robots only have cameras in their head. Problem is, when the robot reaches for something, its own arm blocks the view. Figure-03 solved this by putting cameras in the palms. So even when the hand is wrapped around an object, the robot can still see what its fingers are doing. That's how it managed to rotate the remote and find the right button. It could literally watch its own grip in real time. On top of that, the fingertips have tactile sensors sensitive enough to detect forces as light as a few grams. That's roughly the weight of a paperclip. This means the robot can feel how tightly it's holding something, which is critical when you're switching between a rigid remote control and a soft towel in the same sequence. All of this sensor data feeds directly into Helix-02, which fuses it together and makes split-second decisions about force, grip, and movement. No separate processing pipelines. No handcrafted rules. Just data in, actions out. Look, we're not at the point where you can buy a robot butler. But what Figure is showing here is that the gap is closing fast. This isn't a scripted demo in a controlled lab. It's a robot doing real household tasks, autonomously, in a realistic environment. And each demo they release pushes the bar further. The kitchen demo was impressive. This living room demo is a clear step beyond that. And if the pace of these updates is anything to go by, the next one is going to be even harder to dismiss. A robot dog that can carry 143 pounds, sprint at 11 miles per hour, climb steep hills, and even rally a badminton shuttle in midair. Unitree's new AS2 is not built for lab floors. It is built for the real world. From hiking trails to rough outdoor terrain, this AI-powered quadruped is designed to move with speed, strength, and surprising agility. At just 40 pounds, the AS2 can handle a standing load of up to 143 pounds. That means a robot lighter than many travel suitcases can support the weight of a grown adult. This strength comes from powerful joints that produce up to 90 newton-meters of torque, giving the legs enough force to push off the ground, recover from slips, and stay balanced under pressure. The design focuses on keeping the body light while delivering serious output, so the robot stays quick without sacrificing stability. The AS2 is positioned as a hiking companion rather than a warehouse machine. It can climb stairs about 10 inches high, handle slopes up to 40 degrees, and jump onto platforms nearly 20 inches tall. These abilities allow it to move across rocky paths, forest trails, and uneven ground without constant recalibration. A bionic AI movement model helps adjust posture in real time, improving balance and foot placement. The result is steady forward motion on gravel, dirt, or wet surfaces instead of stiff, hesitant steps. Speed is another major advantage. Under ideal conditions, the AS2 reaches up to 11 miles per hour. Standard versions move comfortably between roughly 6 and 8 miles per hour. That pace allows the robot to keep up during outdoor walks or fieldwork instead of falling behind. Many robot dogs in the market move slower and carry lighter loads, but the AS2 combines both quick movement and real payload capacity in one compact frame. Would you actually trust a robot like the AS2 to hike beside you through rough terrain at 11 MPH while carrying 143 pounds? Is this the future of outdoor robotics, or does it still feel too early? Comment below and share your honest thoughts. Endurance supports that mobility. Higher end versions use a large battery that delivers over 4 hours of runtime when not carrying weight and can cover more than 12 miles in a single session. Even while carrying about 33 pounds, the robot continues operating for more than two and a half hours. This range makes the AS2 practical for long hikes, outdoor inspections, and extended work shifts. A smaller battery option is available for lighter use, and fast charging is supported on select models. Outdoor durability is built into the design. With an IP54 rating, the AS2 resists dust and splashing water, allowing movement in light rain without shutting down. It operates in temperatures ranging from 4 degrees Fahrenheit to over 120 degrees Fahrenheit. That wide range prepares the robot for construction sites, industrial zones, and changing weather conditions. And if you're interested in how machines like this are reshaping outdoor robotics, subscribe for more detailed coverage. The intelligence system ties everything together. The AS2 runs on a high-performance onboard processor and uses LiDAR sensors to scan and understand its surroundings. The upgraded ISS 3.0 intelligent following system allows stable side-by-side tracking with strong positioning accuracy, even on uneven terrain. Built-in connectivity features such as Wi-Fi, Bluetooth, camera systems, and expansion ports support further development. An optional 7-axis robotic arm can be added, allowing the robot not only to carry gear, but also to interact with objects in its environment. One of the most eye-catching demonstrations shows the AS2 playing badminton. The robot tracks the shuttlecock in real time, shifts its weight with quick sidesteps, and swings with controlled balance. Instead of moving in a rigid, pre-programmed way, the motion looks reactive and coordinated, showing how AI, sensing, and mechanical power work together under fast conditions. When a robot can react to a flying shuttlecock in real time, how far are we from machines matching human reflexes? Drop your take below. Unitry has not revealed pricing, and real-world performance may vary from laboratory testing. The company advises users to maintain distance during operation because of the robot's dynamic power. Even so, the AS2 clearly raises the standard for speed, payload, and outdoor mobility. And with robotics moving at this pace, the next breakthrough may already be here.