NVIDIA Releases Newton 1.0 Physics Engine for Industrial Robot Training

NVIDIA has announced the official release of Newton 1.0, an open-source physics simulation engine focused on training robots for contact-rich manipulation and locomotion tasks in industrial environments.

Newton 1.0 exits its beta phase, originally introduced in September 2025, and is now a stable, production-ready platform already seeing uptake in manufacturing for assembly automation. The engine is developed in partnership with Google DeepMind and Disney Research and operates under the Linux Foundation. It is built on NVIDIA Warp, a high-performance computing framework, and OpenUSD (Universal Scene Description), an open standard for 3D data interchange.

The Newton engine incorporates multiple rigid-body solvers, notably MuJoCo Warp and Kamino. According to NVIDIA, MuJoCo Warp, when run on the company's RTX PRO 6000 Blackwell GPU, achieves speeds up to 252 times faster than MJX for locomotion tasks and up to 475 times faster for manipulation tasks. This performance leap is aimed at accelerating reinforcement learning—a type of machine learning where robots learn optimal behavior through trial and error—in industrial settings.

In addition to its rigid-body solvers, Newton offers a Vertex Block Descent solver, which enables simulation of deformable materials such as cables, cloth, and volumetric components. The platform also features a signed distance field collision library and supports hydroelastic contact modeling, further expanding its capabilities for simulating complex physical interactions critical to robotics research and deployment.

Newton is natively integrated with Isaac Lab 3.0, NVIDIA's open-source robot learning framework, which became available in early access alongside Newton 1.0. Partnerships with organizations like Toyota Research Institute are supporting further development of advanced solvers and contact modeling capabilities.

Industry actors are already leveraging Newton’s capabilities. Skild AI utilizes Newton with Isaac Lab to train reinforcement learning policies for tasks such as GPU rack assembly, focusing on precise connector insertion, board placement, and fastening. In the consumer electronics sector, Samsung is working with Lightwheel to apply Newton's deformable simulation capabilities to cable manipulation for refrigerator assembly lines.

At the recent NVIDIA GTC event, the company also announced updates to its broader robotics ecosystem, including the introduction of Cosmos 3, new versions of the GR00T large language model for robots (GR00T N1.7 and a preview of GR00T N2), and revealed that the Isaac and Omniverse frameworks now support over 2 million robots worldwide.

Newton 1.0 is available as open-source software, supporting the growing demand for realistic robot training environments that enable rapid development and deployment in industrial settings.

Source: dataconomy.com