Disney has solved one of the most costly problems in bipedal robotics: falling. A new system based on reinforcement learning1 It trains robots to choose safe landing positions that protect their sensors, batteries, and joints. The method uses 4.000 simultaneous virtual simulations to teach robots to roll, land, and absorb impacts instead of resisting them. Published on arXiv By November 2025, the study shows 16-kg robots falling from lateral thrusts at 2 m/s with controlled landings. No more rigid joints or random movements: every fall becomes a managed, calculated, and safe event.
When falling becomes a 35-kilo problem
Let's leave aside for a moment the robots dancing on conference stages. Those are choreography, not physics. The real problem comes when gravity decides it's its turn. A biped robot that loses its balance turns into a 16-kg projectile falling to the ground with stiff joints, exposed sensors, and no defenses. The result? Cracked shells, damaged electronics, and compromised batteries. And repair bills that hurt.
Disney Research A team of university engineers have spent years teaching robots how to manage falls. Not how to avoid them (that's another topic), but how to land in the least painful way possible. The system is called "Robot Crash Course” and it does exactly what it promises: it trains robots to fall better. A bit like a parkour course, but for machines that weigh as much as Buddy, my golden retriever: to be precise, a medium-sized dog.
AIDOL and the black cloth of shame
Before we talk about how a Disney robot would solve the problem, it's worth remembering what happens when no one solves it. Moscow, November 11, 2025. Stage, lights, Rocky theme playing in the background (which made everything a little more ridiculous, let's face it). Enter AIDOL, the first Russian humanoid robot with artificial intelligence. He raises his right arm in greeting. “Haven't seen the rare one?”2,” we would joke in Naples. The rest is history: AIDOL loses his balance. AIDOL falls face down. The assistants run with a black sheet to cover him from the sight of journalists and guests. A Buster Keaton-esque skit and a subtle satisfaction for us humans in seeing that these technological bogeymen, destined according to popular belief to replace us, still walk as if they have severe diarrhea and frequently spill their guts on the ground.
According to the CEO Vladimir Vitukhin, it was a calibration error in the balance sensors. A “useful accident,” he jokingly called it. Helpful or not, the video went viral in three hours and proved a very simple thing: Falling is the most critical moment in the life of a biped robot. It doesn't matter how sophisticated your control system is if you don't have a plan B for when gravity wins.
The Disney system doesn't prevent a robot from falling like AIDOL's. It manages them. The difference is that a Disney robot trained with this method doesn't end up facedown with its joints locked. It folds up, protects its head, and adopts a final pose that distributes the impact, like a sort of stuntman. And then it gets back up.
Disney Robots: 24.000 Virtual Falls to Learn How to Suffer Less
The heart of the system, I said, is the reinforcement learningThousands of digital robots fall into a simulator, studying what works and what doesn't, and accumulating data. Each fall generates points if the robot reduces the impact or protects critical areas. Points are deducted if the movements become chaotic or out of control. The system has tested 24.000 stable poses, launching virtual models from various heights (always consistent, not ten stories high). Ten of the final poses came from “artists” who created creative positions: crouches, wide rolls, dramatic landings.
The training lasted two days on powerful GPUs. 4.000 virtual robots They fell simultaneously. A small neural network processed joint angles, body orientation, and motion data fifty times a second. The method is called proximal policy optimization and adapts the robot's behavior step by step without sudden jumps. The simulator reduced contact pressure and set different sensitivity levels for each body part: the legs remain soft, while the head requires more protection.
From simulator to lab: testing the real Disney robot
After virtual training, the system was installed in a real Disney robot. It weighed 16 kilograms. It had spring-loaded legs and mechanical arms. A motion capture system tracked every movement and sent updates to the controller in real time. Tests showed that the robot could handle falls from lateral thrusts of 2 meters per second (about 6,5 feet per second) and forward slides with rotating hips. The speeds were randomized in each episode, so the robot never learned a fixed path.
The robot doesn't resist the impact: it embraces it. It rolls, positions itself, and protects its batteries and sensors. Its final position depends on the direction of the fall and its speed. Sometimes it crouches tightly. Sometimes it opens wide. Sometimes it chooses something in between. As I was saying in this article, robotics in 2025 is moving toward more adaptive and intelligent systems. This study confirms it: robots don't just have to walk well. They must also (above all, one might say) fall well.
The most interesting fact? The system works with different robots. The policy is hardware agnostic: it can be transferred to other bipedal platforms without retraining everything from scratch. This means that the method is not tied to a single model, but applies to any machine with two legs and compatible joints.
Why teaching how to fall matters more than we think
The bottom line is simple. Bipedal robots are becoming more common. Some he already put them on pre-sale. Factories, warehouses, hospitals. They walk among us, transporting objects, performing tasks. But as long as falling means breaking, their use remains limited. A robot that can fall without being damaged is a robot that can work longer, in more environments, with less human supervision. It's not a question of elegance: it's pure economy.
Disney Research has chosen a pragmatic approach. It doesn't sell the fall as a spectacle. It treats it as an engineering problem with a measurable solution. The robot falls, protects its critical parts, gets up, and continues. No black tarps, no Rocky theme, no awkwardness. Just applied physics and reinforcement learning that works.
All it takes is a little learning and the robot goes down, if you'll pardon the pun "cringe" used by a dad of an eight-year-old. The important thing, obviously, is that it gets back up in one piece.
- Reinforcement learning is an artificial intelligence technique in which an agent learns to make decisions by interacting with an environment, receiving rewards for correct actions and penalties for incorrect ones. Thus, as happens in our everyday learning through trial and error, the agent improves its behavior to achieve goals over time. It is used to teach machines to solve complex problems where they must make a series of choices to achieve the best possible outcome. ↩︎
- “He didn't see the step.” It's sometimes said, even when there's no step to watch out for, just to pitifully emphasize an unwary fall. Follow me for more unsolicited Neapolitan lessons in the footnotes. ↩︎
