Newton Project Advances Open Robotics Simulation

The Linux Foundation is a non-profit organization that governs and supports community-focused, open-source projects for software and hardware developers and pioneers worldwide. The organization encourages development through training, partnerships, open-source development, and code sharing. The foundation recently announced its governance of the jointly developed open source, GPU-enhanced, scalable physics simulation engine, Newton. The engine is built upon the developer framework, NVIDIA Wrap, and the Open USD ecosystem.

The Newton Physics project is available to developers worldwide, providing GPU-accelerated simulations for driving forward application development in general-purpose robotics. Image used courtesy of NVIDIA Developer

Newton: An Overview

Newton’s foundational architecture is ModelBuilder, which creates models from imported assets and application parameters. Such parameters could include fixed structural properties of robots (including mass, inertia tensor, positioning, velocity states, kinematic and dynamic flags, damping, collision/geometry, etc.), actuation, material properties, sensors, etc.

The ModelBuilder’s resulting output model encapsulates the above parameters, physical structure, and organization. The State element of Newton’s architecture captures all the mutable parameters that change with time or each simulation phase. The Solver component of the architecture integrates physics to enhance the simulation. Lastly, the Viewer component aids in visualizing the simulation, whether online or offline.

GPU-Enhanced Code and Modular Solver Setup

The NVIDIA Warp framework allows developers to generate low-level code using Python syntax (expressed as Warp Kernels) that run in parallel on the GPU to speed up simulation by orders of magnitude. The ability to add many small, parallel GPU programs in the form of Warp Kernels facilitates advanced AI training with the capacity to run thousands of simulations at one time and as many (or tens of thousands more per second).

The research-friendly and flexible aspect of Newton’s design is the ability to support multiple solvers (or many unique algorithms) for divining how certain robotic hardware moves under certain control commands.

Depending on their application preferences and requirements, users can choose between solvers (such as Euler, VBD, XPBD, MuJoCo, and Featherstone). For example, developers might choose Featherstone to facilitate the movement of rigid-body robotics systems (such as robot arms and humanoids), or XPBD for soft robotics and real-time applications. Developers might choose Euler for more basic simulations required for testing and training. Developers can combine solvers (or hybridize their pipeline) to handle different elements of robotic movement and control for a given application. As the Solver component of the Newton architecture is separate from the Model, developers can swap solvers without needing to rebuild the robot model.

The in and outs of the Open USD framework: a powerful tool for simulation workflows. Video used courtesy of NVIDIA Developer

The foundational Open USD element of Newton allows developers to standardize model imports, retaining important physical properties, geometry, and other robotic/environmental parameters. This guarantees uniform physical modelling across platforms, streamlines workflow integration, and renders Newton compatible with numerous other simulation, graphics, and robotics programs.

Purpose

The Newton project forges a path towards enhanced robotics innovation at a global scale, providing the tools and open-source, community-focused practices of the Linux Foundation to democratize and speed up the development of general-purpose robots. Speed of development is not the only key output here, but creating affordable, simulation-tested robotics solutions for optimum performance in real-world settings.