New Lightbot MASTERED Parkour… This AI Robot Is INSANE!

We've already seen Unitrix-G one-run, jump, fight, and pull-off movements that looked almost impossible for a robot this small. Now an even tinier humanoid called Lightbot Zero is taking that idea in a completely different direction, using AI to climb, vault, balance, and basically turn itself into a parkour expert. And what makes it interesting is that the robot is not simply copying a fixed routine, it is learning how to deal with obstacles it has never encountered before. But while humanoid robots keep getting crazier, China's robotics race is starting to expand far beyond just building stronger bodies. We have already seen the raw power of Engine AI's T-800 and some incredibly realistic humanoids, but another company is focusing on something much more practical. X-Square Robot is building an AI system designed to attack one of the biggest headaches in logistics, sorting messy, unpredictable packages at high speed. And if you remember Figure-03's massive live package sorting demonstration, X-Square has now put its own system through a very similar test. The difference is that X-Square does not even need a humanoid body to show just how far robot intelligence is starting to go. So, let's start with the first update. Light Origins has introduced Lightbot Zero, a small humanoid robot built to do parkour. It stands about 90 centimeters tall and weighs under 19 kilograms, which puts it closer to the size of a child than the heavy humanoids we usually cover. Watch what it actually does. It runs at a waist-high box and vaults straight over it with one hand planted on the surface. It climbs onto a raised platform by pulling with its arms and pushing with its legs, then steps back down without hesitating. It crosses a plank barely wider than its own foot. It picks its way across scattered stepping stones and walks up outdoor stairs with curbs. At one point, it clears a gymnastics pommel horse it had never encountered before. All of this happens with a single camera on its chest. There is no motion capture in the room and no map of the obstacle loaded in advance. You simply tell the robot to move forward and it decides for itself whether that means walking, climbing, or vaulting. Getting a robot to move like this has always been the hard part. The usual approach is to record a human doing the move and copy it onto the robot. The trouble is that a video captures the person, but not the obstacle they were pushing against. Copy that motion across, and the robot's hands press down on empty air while its knees pass through the box that should be holding it up. The movement survives and the meaning is lost. Light Origins fixed this by recording one short human clip for each move. A physics simulation then rebuilds the missing half, placing the obstacle where the hands need to land and reshaping the motion until the robot can genuinely perform it. Then comes the clever part. Every time the robot succeeds, that attempt becomes the new lesson and the obstacle rises a few centimeters. One 45 centimeter climb grew into a robot that handles 75, which is over 80 percent of its own height. Picture an adult hauling themselves onto something chest high. They also gave it a short memory, and that turned out to matter enormously. During a vault, the chest camera loses sight of the box at the exact moment the hands come down on it. The robot is blind through the most important instant, so it simply remembers what it saw a second earlier. Strip that memory away, and its success rate collapses. The results back it up. At 70 centimeters, the robot clears the obstacle nine times out of ten. Teaching it to flow between walking and parkour lifted its course completion from roughly a third to almost every attempt. None of this needed real-world practice. Everything was learned in simulation and worked on the physical robot straight away. Parkour was only the test here because the same approach already carried over to a robot learning to pick things up. Before we go further, if you enjoy humanoid updates, make sure to subscribe. You'll get more updates like this and follow the biggest breakthroughs as AI and robots continue to evolve. Picture a robotics company going live in front of the whole internet. Robots at a workstation sorting parcel after parcel, the package count climbing on screen. No edits, no cuts, nothing hidden. And the point is not just the hardware, it is the AI brain running the hardware. If figure AI came to mind, you are right. Figure did exactly that, streaming past 200 hours with a rotating team of figure 03 humanoids that swapped in a fresh unit whenever a battery ran low. By the end that run had sorted close to a quarter of a million packages. It settled into an average of around 1,248 parcels an hour. But figure is not the only company that ran this play. A robotics company in China called X-Square Robot decided to prove a point of its own and went live. Two things were on camera at the same time. One was a pair of robotic arms sorting parcels at a workstation. The other was the AI brain telling those arms what to do, a model the company calls wall-b. So what were those arms actually up against? The job is called parcel induction and it is the ugliest stage of warehouse sorting. Packages turn up dumped in overlapping piles where soft bags slump against rigid boxes and labels point every which way. Every time the arm pulls one item out, everything left behind shifts around, so the pile it sees is never quite the pile it saw a second earlier. For each parcel, the arms had to find one item in that tangle and lift it cleanly, turn it in the air until the barcode faced the scanner, then send it on its way. Nothing was edited, so there was no chance to hide a mistake or quietly start over. The arms simply worked for a full hour with nobody helping them. They finished with 1,816 parcels sorted. The success rate held above 98% the whole way through. That works out to about one package every two seconds. Here is the interesting part. Before going live, X-Square set itself a target that landed almost exactly on figure's hourly average. The company never explained why, so nobody can call it a direct challenge. X-Square was chasing something bigger anyway. It wanted to show you do not need legs, or a torso, or a human-shaped body to be brilliant at this job. And then came the moment that made the whole stream worth watching. A parcel slipped and started sliding off the workstation. The gripper caught it mid-fall and carried on without losing rhythm. That save is where the brain shows itself. Ordinary factory automation is exactly what falls apart in a moment like that because it follows a planned motion and has nothing to offer once the world moves underneath it. But their robot AI brain, Wall-B, works differently. Most robots handle seeing, understanding and instruction, and moving as three separate pieces that get trained apart and bolted together later. Wall-B learns all of it inside one connected model, similar to figure's Helix-02 AI brain. The robot understands what it is looking at and what is going to happen once it reaches in. It has a working feel for weight and friction and objects knocking into each other, which is why it can tell a bag is about to fold or a parcel is about to slide before either one actually does. X-Square trained it on messy footage from real workspaces instead of tidy laboratory runs, and you can see that choice pay off every time the pile misbehaves. The company has taken the idea further with a project called Host, where the robot picks up a A single demonstration under 30 seconds is enough. They tried it on 50 tasks the robot had never seen before, and it got them right more than half the time, with no retraining at all. That tells you what X-Square is really building. The body stays locked to one job while the intelligence sitting on top stays easy to teach. So the arms are simple on purpose, and the brain is anything but. X-Square is betting that the smartest part of a robot does not need to look human to matter. After an unbroken hour in front of a live audience, that bet is looking a lot stronger than it did a week ago. So from a tiny robot learning parkour to an AI system sorting thousands of messy packages, robot progress is clearly moving in more than one direction. Let me know which update impressed you more, and subscribe for more AI and robotics news like this.