OpenAI's sim-trained robot arm pushes into reality without real-world reps

The oldest problem in robotics might have just found its simplest solution. For decades, engineers have grappled with the ‘reality gap’—the infuriating tendency of a robot that performs flawlessly in a pristine simulation to flail uselessly the moment it touches real hardware. A new paper from OpenAI demonstrates a method so straightforward it almost feels like cheating: instead of trying to build a perfect simulation, they deliberately broke it. By randomizing the simulator’s dynamics during training—constantly changing things like mass, friction, and joint behavior—they forced the AI to develop policies that don’t rely on any single, brittle model of physics. You’re not teaching the robot how to push a puck across one specific table; you’re teaching it how to push a puck across any table, including ones with physics it has never technically seen before. The result is a control policy that adapts on the fly, compensating for the messy, unpredictable nature of physical reality without ever having touched it. And they didn’t just simulate this. They deployed the policy on a real robotic arm tasked with moving an object to a target location. The robot, trained exclusively inside a computer, maintained near-identical performance when confronted with actual hardware, reliably finishing the task from randomized starting positions. That’s a big deal. Traditional sim-to-real transfer often requires at least some fine-tuning on the physical robot, a time-consuming and potentially destructive process. This approach skipped that step entirely. The paper explores the sensitivity of their design choices, revealing that the resulting policies show a surprising tolerance for calibration error—meaning you don’t need a million-dollar motion-capture setup to make this work. The implications extend beyond a single pushing task. If randomization proves broadly applicable, the cost and speed of training sophisticated robotic skills could plummet, shifting the bottleneck from physical experimentation to sheer compute power.