By Bobby Carlton
Here’s something that flew under the radar for most people outside the robot simulation weeds. NVIDIA, Google DeepMind, and Disney Research spent the last year building a physics engine together. Not three competing engines. One. It’s called Newton, it went stable earlier this year, and it’s free.
If you don’t live in simulation land that might sound like a footnote. It isn’t. It’s one of the bigger shifts in how robots get built, and it changes the math on a lot of projects that used to be too expensive to bother with.
Let me back up and explain why anyone should care.
Every robot you’ve ever seen do something useful learned to do it somewhere. The good ones didn’t learn on the factory floor. They learned in a simulator, where you can run ten thousand copies of the same robot trying the same task overnight, let them fail a million times, and keep only what worked. Then you take that trained brain and drop it into the real machine. That’s the whole game. Train in the virtual world, deploy in the real one.
The catch has always been the gap between those two worlds. Simulators were never quite right. The friction was a little off. A cable behaved like a stiff rod instead of a floppy wire. A robot hand would grip perfectly in sim and then fumble a real part because the contact physics lied to it. People in the field call this the sim-to-real gap, and closing it has eaten more engineering hours than anyone wants to admit.
Newton goes straight at that problem. It runs on the GPU, which means it’s fast in a way that matters. We’re talking hundreds of times faster than the older tools for the hard stuff like locomotion and manipulation. It handles the things that used to break simulations. Cables. Cloth. Deformable materials. The messy contact-rich moments where a gripper actually touches a part. Those aren’t edge cases. That’s most of real work.
And because three of the heaviest hitters in the space built it together and handed it to the Linux Foundation to manage, it’s not a walled garden. It plugs into MuJoCo, it plugs into Isaac Lab, it speaks OpenUSD so your assets move around without a fight. Disney’s already using it for the expressive droids that walk around their parks. Manufacturers are using it for assembly automation right now.
So that’s the news. Here’s why it matters to you if you run a warehouse, a farm, a plant, or a fleet of anything.
The cost of trying just dropped. For years, if you wanted to know whether a robot could do a specific job in your specific building, the honest answer was “build it and find out,” and that was expensive enough that most good ideas died in a meeting. Now you can answer that question in simulation first. Will this mobile manipulator actually navigate my loading dock? Can this arm handle the weird shape of my product without crushing it? What happens when the floor is wet, the lighting is bad, the pallet is loaded wrong? You can ask all of that virtually, with physics you can finally trust, before anyone signs a purchase order.
That’s the part that excites me. Better tools at the bottom of the stack mean more shots on goal at the top. A simulation that used to take a specialized team and a pile of money now takes a smaller team and a lot less of it. The questions that were too expensive to ask are suddenly worth asking.
We’ve been building in this world for a long time, and the honest truth is that a free, accurate, fast physics engine doesn’t put simulation shops out of business. It does the opposite. The hard part was never the physics solver. The hard part is everything around it. Getting your actual facility into the simulator as a faithful digital twin. Building sim-ready assets of your real equipment and your real products. Generating the synthetic data to train a model on tasks it’ll never see enough of in the real world. Wiring it all into a pipeline that produces something you can act on instead of a pretty demo.
Newton makes the engine room better. Somebody still has to build the ship.
If you’ve got a process you’ve been curious about but never thought was worth simulating, the calculus may have changed this year. The tools got cheaper and better at the same time, which doesn’t happen often. Might be worth a conversation.
Why Teleoperation Data Can't Be Faked. By Bobby Carlton Teleoperation data is the quiet force…
Good chance you’re hearing a lot of conversation around a technology called Gaussian splatting (or…
While generative AI is mainly known for its ability to create text-based models, it is…
Machine Learning Meets Computer Vision: How Synthetic Data and Simulation Are Powering the Future of…
Data is the unsung hero behind every robotic revolution, it’s the raw fuel that turns…
In a recent CNET article, it was announced that Amazon is deploying humanoid robots from…