How does Atlas learn? | Inside the Lab | Boston Dynamics

### TL;DR Boston Dynamics’ new Atlas humanoid learns complex manipulation—like lifting heavy refrigerators—by training in simulation, enabling rapid iteration from design to real-world execution with minimal gap. --- ## Inside the Lab: How Atlas Learns ### A New Era for Atlas “I’ve been at Boston Dynamics for 15 years, leading the controls team for almost all of that time. We’ve gone through many cycles and turns, and I think this phase is the most exciting.” The speaker, a control team lead, joined Boston Dynamics six weeks before filming. His first task? The refrigerator project—originally a year-end sprint goal that turned into a two-week push. “The new Atlas robot is really exciting. When we started designing it, the goal was a low-cost, easy-to-manufacture, reliable, and simple robot. We reduced complexity as much as possible.” Despite the simplification, Atlas retains the traits that make Boston Dynamics robots iconic: “It’s very strong, very flexible, very fast, and easy to control.” ### The Refrigerator Challenge “Atlas needs to be a tool. A tool that works like a human—interacting with objects of all shapes and sizes. And because Atlas is a beast, we knew it could lift heavy things. We wanted to jump straight into the deep end.” The empty refrigerator weighs 50 pounds, which was the primary test weight for filming. In the lab, the team successfully lifted 100 pounds to stress-test the robot. “No other humanoid that I know of has moved objects as big and heavy as a refrigerator.” The key innovation: whole-body manipulation. “Not just grabbing with hands and moving it like this—but really using the entire body, like a straight-arm lift with the whole weight pressed into the body. That was the idea. I’ve always wanted to do that with a robot. Recently, everything came together: a new powerful robot capable of doing it, and a team with cool new tools to program these behaviors.” ### Learning Through Simulation Simulation is the core of Atlas’s learning process. “We design behaviors in simulation, and then almost always they transfer perfectly to the real robot. It plays very successfully.” The workflow is streamlined: “I can train a strategy today, get the baked strategy tomorrow, and within an hour my strategy is on the robot. I test, collect data, see how it works, and iterate. The whole pipeline is seamless. The simulation-to-reality gap is very small. The robot is the robot. Focusing on that without worrying about other factors is really fun.” For the refrigerator task, the team first created an animation reference showing how the robot should move the fridge. Then the robot practiced the exact same motion in simulation for millions of hours. “We encourage it to move the fridge as instructed, not drop it, etc. Once the reward tuning is correct and the physics of the drag is accurate in simulation, running it on the real robot works perfectly.” ### Rapid Iteration Workflow A robust process, already established, makes testing easy. “There’s a whole system—a packaged platform for testing new strategies. Many people work behind the scenes.” The six-week experience demonstrated how quickly ideas turn into actions. “In these six weeks, I noticed testing things has become fairly easy. The entire system is packaged so we can focus on the robot.” ### Pushing the Limits “We haven’t seen Atlas’s limit yet. We are still pushing the boundaries of how big and heavy an object it can carry. And we’ll keep pushing.” The backflip still works: “Three, two, one. Hey! Like a champion.” ### What’s Next? “Now, taking the robot out of the lab has never been easier. We are very excited about that. Right now, our only limit is imagination.” --- **Source:** [How does Atlas learn? | Inside the Lab | Boston Dynamics](https://www.youtube.com/watch?v=xKK5ze3FukQ)