What is robotic
simulation?
A robotic simulator is a computer program that builds a virtual copy of a robot and its surroundings. Inside it, engineers can test, break and train the robot thousands of times without touching real hardware.
The four parts
The physics engine. The part that computes movement. It works out, many times per second, where every part of the robot will be next, how hard its feet push on the floor and whether it falls over. In NVIDIA Isaac Sim, the physics engine is called PhysX, with a newer one called Newton being introduced.[1] More in What is a physics engine?
The renderer. The part that draws the picture. It produces the images a camera on the robot would see. If the pictures are realistic enough, they can be used to train the robot's vision, and each pixel can be labelled automatically, which is called synthetic data.[1]
The robot description. A file that lists the robot's parts, the joints between them, their weights and their 3D shapes. Common file formats are URDF, MJCF and USD; each simulator prefers one of them.[1]
Virtual sensors. Simulated cameras, laser scanners and motion sensors that produce the same kind of data as the real devices, so the robot's software cannot tell the difference.[1]
Why train a robot in software
A 2025 survey of the field gives the reason in one line: collecting training data on real robots is "costly, time-consuming, and unsafe".[5] A real robot that falls can break. A real robot learns one attempt at a time. A simulated robot can fall a million times for free.
The second reason is scale. In March 2025, NVIDIA reported producing 780,000 synthetic robot movements in 11 hours, which it compared to about nine months of human demonstrations.[2] This is what makes reinforcement learning practical: the robot practises in thousands of parallel copies of the world at once.
What is changing now
Physics engines are being rebuilt for graphics cards and opened up. Newton, developed by NVIDIA, Google DeepMind and Disney Research, was handed to the Linux Foundation in September 2025 so that no single company owns it.[3] Simulators are also getting broader: Genesis, an open-source project started in December 2024, can simulate rigid parts, soft materials and liquids in one scene.[4]
What most simulators still leave out is the physics underneath the mechanics: how hot a motor gets, how much battery is left, how strongly a magnet or a suction cup holds. That gap is what Invisible Force works on. See What is the sim-to-real gap?
References
- NVIDIA. Isaac Sim documentation: overview. September 2026.
- NVIDIA Newsroom. NVIDIA Isaac GR00T N1 press release, announcing the Newton physics engine. March 2025.
- Linux Foundation. Linux Foundation announces contribution of Newton by Disney Research, Google DeepMind and NVIDIA. September 2025.
- Genesis Embodied AI. Genesis: a generative and universal physics engine for robotics (repository README). December 2024, updated 2026.
- Wong, Kang, Bai and Zhang, arXiv 2505.01458. A Survey of Robotic Navigation and Manipulation with Physics Simulators in the Era of Embodied AI. May 2025, revised June 2026.
Every link was opened and checked in September 2026. Links open in a new tab.