NVIDIA Newton

Scroll sideways for more →

Newton is an open-source, GPU-accelerated physics engine built for robotics simulation and robot learning, co-developed by NVIDIA, Google DeepMind, and Disney Research and stewarded by the Linux Foundation. It is built on NVIDIA Warp and OpenUSD, is differentiable, and is designed to give robots (especially humanoid robots and robotic characters) fast, high-fidelity simulation for sim-to-real transfer and reinforcement and imitation learning.^[1][2][4] Newton is not a single physics solver but an extensible framework that hosts multiple solvers behind a common interface, and it is meant to plug into existing robot-learning stacks rather than replace them.^[6][7]

What Newton is, and why robots need a physics engine

A physics engine is software that simulates how bodies move and interact: how a joint rotates under torque, how a foot contacts the ground, how friction and inertia play out over thousands of tiny time steps. For robotics, a physics engine is the substitute for reality during training. Modern robot control policies, particularly for legged locomotion and dexterous manipulation, are learned with reinforcement learning and imitation learning, which require enormous numbers of trial-and-error episodes. Running those episodes on physical hardware is slow, expensive, and unsafe, so developers run them in simulation and then transfer the learned policy to the real robot. The accuracy of that simulation determines how well the policy survives contact with the real world, the problem known as the sim-to-real gap.^[2][4]

Newton exists to make that simulation loop both faster and more accurate. NVIDIA describes it as a way to cut simulations "from days to minutes" by running them on GPUs, and to close the sim-to-real gap with accurate contact, friction, and actuator modeling.^[1] Because it is differentiable, Newton can also propagate gradients through the simulation itself, which supports gradient-based training, design optimization, and system identification rather than only black-box trial and error.^[1][2] In NVIDIA's framing of "three computers" for embodied AI, Newton is the one that "simulates the body" of a robot, sitting between the Isaac GR00T robot foundation models (the "brains") and NVIDIA Omniverse (the "training ground").^[5]

History and the GTC 2025 announcement

NVIDIA unveiled Newton on March 18, 2025 during Jensen Huang's keynote at GTC 2025, announcing it as an open-source physics engine co-developed with Google DeepMind and Disney Research.^[2][3] The reveal was paired with the debut of the Isaac GR00T N1 humanoid foundation model and a set of simulation libraries, positioning Newton as the physics layer of a larger physical AI push.^[3] Two of Disney's Star Wars-inspired BDX droids walked onstage with Huang during the keynote as the first public showcase of what the collaboration was for.^[3][12][17]

At announcement, NVIDIA said it aimed "to make the first version available later this year," meaning 2025, rather than shipping code that day.^[2] The project matured on a public timeline:

Scroll sideways for more →

The distinction matters: the September 2025 event was a beta and the point at which the project moved to open governance, while "Newton 1.0" is the later general-availability release from GTC 2026.^[4][5][6]

The NVIDIA, Google DeepMind, and Disney collaboration

Newton is a genuinely three-party project, and each partner brought a distinct piece.^[2][4]

• NVIDIA contributed the compute foundation: Warp (its GPU simulation framework), OpenUSD integration, and the connections to Omniverse, Isaac Sim, and Isaac Lab.^[1][2]
• Google DeepMind contributed MuJoCo expertise. DeepMind, which maintains the widely used MuJoCo simulator, introduced MuJoCo Warp (also written MJWarp), a GPU-accelerated implementation of MuJoCo that became one of Newton's core solvers. At announcement DeepMind and NVIDIA reported that MuJoCo-Warp delivered "more than a 70x acceleration for humanoid simulations and a 100x speedup for in-hand manipulation tasks."^[2][14]
• Disney Research was both a co-developer and the first adopter. Newton's architecture let Disney build Kamino, a GPU-accelerated simulator that, in Disney's words, "unlocks the potential of reinforcement learning" for robotic systems of unusual mechanical complexity, such as legged characters with closed-loop linkages.^[4][6] Disney's stated plan is to use Newton for its next-generation entertainment robots, including the Star Wars-inspired BDX droids rolling out to its parks.^[3][14]

These vendor multipliers ("more than 70x," "100x") are the developers' own benchmark claims rather than independently verified figures, and the exact numbers changed as the software matured (see Architecture below).^[2][6]

Linux Foundation governance

On September 29, 2025, at the Conference on Robot Learning (CoRL) in Seoul, Disney Research, Google DeepMind, and NVIDIA contributed Newton to the Linux Foundation "to accelerate open robot learning."^[4][5] Moving the project under a neutral foundation is meant to prevent any single company from controlling the de facto standard simulator and to encourage community-driven contribution. Jim Zemlin of the Linux Foundation called it "an important step forward for scaling collaborative robotics simulation," and NVIDIA's Rev Lebaredian framed Newton as bringing together "GPU acceleration, differentiable physics and open standards."^[4]

The Linux Foundation release also named early supporting organizations beyond the three founders, including the Technical University of Munich, Peking University, Lightwheel, and Style3D.^[4] Newton's code is licensed under Apache-2.0 and its documentation under CC-BY-4.0, and development happens in the open at github.com/newton-physics/newton, where the project is described as "community-built and maintained."^[7]

Architecture and how it works

Warp foundation and GPU acceleration

Newton is built on NVIDIA Warp, a Python framework for writing GPU-accelerated simulation and spatial-computing kernels that compile down to CUDA. Warp lets Newton run many simulation environments in parallel on a single GPU, which is what turns days of CPU simulation into minutes and makes large-batch reinforcement learning practical.^[1][2] Warp is also what makes Newton differentiable: it can compute reverse-mode gradients of the simulation so that developers can back-propagate through physics.^[2][7]

OpenUSD as the data model

Newton uses OpenUSD (Universal Scene Description) as the format for describing robots and their environments. OpenUSD's composition engine aggregates the geometry, joints, sensors, and scene data a simulation needs, and it is the same interchange layer used across Omniverse and Isaac Sim, so assets can move between authoring, simulation, and training without conversion.^[1][2]

Differentiability

Differentiable physics is a headline feature. Because gradients can flow through the simulator, Newton supports gradient-based policy training, design optimization (tuning a robot's physical parameters), and system identification (fitting the simulation to match a real robot). NVIDIA says this "accelerates training, design optimization, and system identification" compared with gradient-free methods alone.^[1][2][7]

MuJoCo Warp and Isaac Lab compatibility

Newton is designed to slot into existing robot-learning workflows rather than force a migration. It is compatible with Google DeepMind's MuJoCo and MuJoCo Playground and with NVIDIA Isaac Lab, the open-source unified framework for robot learning.^[1][2] With Newton 1.0, it integrates natively with NVIDIA Isaac Sim 6.0 and Isaac Lab 3.0 as a new physics and camera-sensor backend, meaning teams already using Isaac can swap Newton in as the underlying solver.^[6]

The MuJoCo Warp benchmark numbers were sharpened over time. At the March 2025 announcement, DeepMind and NVIDIA cited "more than 70x" for humanoid simulation and "100x" for in-hand manipulation.^[2] By the Newton 1.0 release in March 2026, NVIDIA reported that MuJoCo Warp running on an NVIDIA RTX PRO 6000 Blackwell GPU sped up MJX (MuJoCo's earlier GPU path) by 252x for locomotion tasks and 475x for manipulation tasks.^[6][15] These are NVIDIA's own hardware-specific benchmarks and should be read as vendor claims, not neutral measurements.

Extensibility and multiple solvers

Rather than a single monolithic solver, Newton is a framework that hosts several, chosen per task. Version 1.0 ships with multiple rigid-body solvers, the two most capable being MuJoCo Warp (from DeepMind, a GPU version of MuJoCo 3.5) and Kamino (from Disney Research, built for mechanisms such as robotic hands and legged systems with closed-loop linkages and passive actuation). It also includes Vertex Block Descent (VBD) for deformable simulation and an implicit Material Point Method (iMPM) for particle and granular simulation. Developers can plug in custom solvers for rich multiphysics scenarios.^[1][6]

Materials and solvers, including the SIGGRAPH 2026 enhancement

Most robotics simulators concentrate on rigid bodies (links, joints, contacts) because that is what covers walking and grasping. Newton extends beyond that to deformables and particles, and this is where its 2026 research work is concentrated.

At SIGGRAPH 2026 (July 20, 2026), NVIDIA presented a new solver that, in NVIDIA's words, "brings hard-to-simulate materials, such as snow, sand and elastic solids, to life inside the NVIDIA Newton physics engine."^[8][16] The solver is the work described in the paper "Mixed Material Point Methods for Stiff Elastoplasticity" by Gilles Daviet, published in ACM Transactions on Graphics for SIGGRAPH 2026. It is a first-party module in Newton that extends the Material Point Method (MPM), a hybrid particle-grid technique well suited to materials that flow, pile, and fracture. The paper demonstrates granular materials (a sand simulation with 49 million particles), snow (including avalanches with fracture propagation), and stiff elastic and elastoplastic solids up to the incompressible limit, and it supports two-way coupling with rigid-body solvers so that grains and soft materials push back on articulated robots and rigid obstacles and vice versa.^[9]

One caution on speed: NVIDIA's SIGGRAPH messaging emphasizes research "that behave[s] realistically and respond[s] in real time" as a theme across its 21 accepted papers, but the specific materials solver is not, on the evidence of its own paper, a real-time method. The paper reports roughly 4 seconds per frame for the large 49-million-particle sand simulation on a single GPU, which is fast for that scale of stiff-material simulation but is offline rather than interactive.^[9][16] The value for robotics is fidelity and coupling, letting a robot be trained against snow, sand, or soft terrain it will actually encounter, not necessarily live interactivity.

Adopters and use cases

Disney entertainment robots: BDX droids and Olaf

Disney is Newton's flagship user. The Star Wars-inspired BDX droids that debuted at GTC 2025 are built on the Disney robotic-character platform that Newton underpins, and Disney has said it plans to bring more such characters "to life in ways the world hasn't seen before."^[3][14][18]

The most visible demonstration came at GTC 2026, where a free-roaming robotic Olaf, the snowman from Disney's Frozen, walked and talked onstage with Jensen Huang. NVIDIA described the character as "driven by NVIDIA's physical AI stack, the Newton physics engine and NVIDIA Omniverse-powered simulation," with an NVIDIA Jetson module as its onboard compute.^[10] Unlike a traditional fixed animatronic, this Olaf balances and locomotes on its own. It was trained with deep reinforcement learning inside Disney's Kamino simulator, one of Newton's rigid-body solvers built on Warp, running many parallel virtual Olafs on a single GPU so the robot could learn to walk and keep its balance in hours of wall-clock training. According to Disney and press coverage, the training ran on the order of 100,000 virtual Olafs over roughly two days on a single consumer GPU, and the physical robot later performed for guests at Disneyland Paris's World of Frozen.^[10][11] The point of the demo was sim-to-real: because Olaf trained inside high-fidelity Newton physics, it could step out of simulation and adapt to a real stage and a moving boat.^[11]

Humanoid and industrial robotics

Beyond entertainment, Newton targets general robot learning for humanoid robots and industrial arms. Newton 1.0 was positioned specifically for contact-rich manipulation and locomotion in industrial settings, and NVIDIA named early industrial users including Skild AI (for GPU rack assembly), Samsung (cable manipulation and refrigerator assembly), and Lightwheel (SimReady asset development).^[6] Because Newton pairs with the Isaac GR00T robot foundation models and Isaac Lab, it fits into pipelines that generate synthetic data, train policies, and deploy them to hardware, and it is complementary to NVIDIA's broader physical-AI stack including NVIDIA Cosmos world models and digital twin workflows.^[5][6]

Newton versus other simulators

Newton is not the only robotics simulator, and it is often complementary to the others rather than a straight replacement (it can act as the physics backend inside Isaac Sim and Isaac Lab, and it reuses DeepMind's MuJoCo as a solver). The table below compares the main options on the axes that matter for robot learning. Claims about competing tools reflect their general design as of mid-2026.

Scroll sideways for more →

Newton's differentiators are that it is GPU-native from the ground up, differentiable, open source under neutral governance, and built as a multi-solver framework rather than a single engine. Its closest peer on capability is MuJoCo, which is unsurprising given that MuJoCo Warp is one of Newton's own solvers; the two are convergent rather than rival.^[2][6][7]

Significance

Newton matters for three reasons. First, consolidation: robotics has long been fragmented across simulators with different physics, formats, and quirks, which made results hard to reproduce and policies hard to transfer. A GPU-native, OpenUSD-based, open-source engine backed by NVIDIA, DeepMind, and Disney, and governed by the Linux Foundation, is a credible candidate for a shared standard.^[4][5] Second, speed and differentiability change what is trainable: batching thousands of environments on one GPU and back-propagating through physics makes reinforcement and gradient-based learning far cheaper, which is a bottleneck for embodied AI.^[1][2] Third, it is a keystone of NVIDIA's physical-AI strategy, the "body" computer alongside GR00T foundation models and Omniverse, aimed at the wave of humanoid robots and general-purpose machines that vendors are racing to ship.^[5][6]

Limitations

Newton is young. It was announced in March 2025, released in beta only in September 2025, and reached its 1.0 general-availability milestone in March 2026, so its production track record is short compared with mature tools such as MuJoCo or Gazebo.^[2][4][6] Its performance headline numbers are vendor benchmarks tied to specific NVIDIA hardware and specific tasks, and should be treated as such rather than as universal speedups.^[2][6] Being GPU-native, it is oriented toward NVIDIA GPUs and the CUDA and Warp stack, which is a practical dependency for teams without that hardware. The advanced material solvers (snow, sand, soft solids) demonstrated at SIGGRAPH 2026 are high-fidelity but not necessarily real-time, so their near-term use is offline data generation and training rather than live interaction.^[9] And as with every simulator, the sim-to-real gap is narrowed but not eliminated: fidelity in contact, friction, and deformable dynamics is exactly where simulators still diverge from reality, which is the problem Newton is built to chip away at rather than one it claims to have solved.^[2][9]

ELI5

Robots learn to walk and grab things by practicing millions of times. Doing that with a real robot would take forever and break a lot of robots, so engineers build a super-accurate video-game version of the world and let the robot practice there instead. Newton is that video-game world for robots. It runs on a graphics chip (a GPU), so it can run thousands of practice robots at once and finish in minutes instead of days. Three big groups built it together: NVIDIA (the chip company), Google DeepMind (an AI lab), and Disney (yes, the movies company, which wanted its robot characters to move for real). Disney used it to teach a walking, talking Olaf snowman to balance on stage, and factories use it to teach robot arms to assemble things. Because Newton is free and open, anyone can use it and improve it, and a neutral group (the Linux Foundation) looks after it so no single company owns it.

See also

• MuJoCo
• NVIDIA Isaac Sim
• Isaac Lab
• Isaac GR00T
• NVIDIA Omniverse
• Robot learning
• Sim-to-real
• Physical AI
• Humanoid robot

References

1. NVIDIA Developer, "Newton Physics Engine," developer.nvidia.com. https://developer.nvidia.com/newton-physics ↩
2. NVIDIA Technical Blog, "Announcing Newton, an Open-Source Physics Engine for Robotics Simulation," March 18, 2025. https://developer.nvidia.com/blog/announcing-newton-an-open-source-physics-engine-for-robotics-simulation/ ↩
3. NVIDIA Newsroom, "NVIDIA Announces Isaac GR00T N1, the World's First Open Humanoid Robot Foundation Model, and Simulation Frameworks to Speed Robot Development," March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks ↩
4. Linux Foundation, "Linux Foundation Announces Contribution of Newton by Disney Research, Google DeepMind and NVIDIA to Accelerate Open Robot Learning," September 29, 2025. https://www.linuxfoundation.org/press/linux-foundation-announces-contribution-of-newton-by-disney-research-google-deepmind-and-nvidia-to-accelerate-open-robot-learning ↩
5. The Robot Report, "NVIDIA launches Newton physics engine and GR00T AI at CoRL 2025," September 29, 2025. https://www.therobotreport.com/nvidia-launches-newton-physics-engine-gr00t-ai-corl-2025/ ↩
6. NVIDIA Technical Blog, "Newton Adds Contact-Rich Manipulation and Locomotion Capabilities for Industrial Robotics," March 2026. https://developer.nvidia.com/blog/newton-adds-contact-rich-manipulation-and-locomotion-capabilities-for-industrial-robotics/ ↩
7. newton-physics/newton, GitHub repository (Apache-2.0 code, CC-BY-4.0 docs; Linux Foundation project). https://github.com/newton-physics/newton ↩
8. NVIDIA Blog, "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI," July 20, 2026. https://blogs.nvidia.com/blog/siggraph-news-2026/ ↩
9. Gilles Daviet (NVIDIA), "Mixed Material Point Methods for Stiff Elastoplasticity," ACM Transactions on Graphics (SIGGRAPH 2026). https://research.nvidia.com/labs/prl/mixed_mpm/ ↩
10. NVIDIA Blog, "NVIDIA GTC 2026: Live Updates on What's Next in AI," March 2026. https://blogs.nvidia.com/blog/gtc-2026-news/ ↩
11. Disney Experiences, "NVIDIA GTC: Walt Disney Imagineering's Olaf Robotic Character Appears," March 2026. https://disneyexperiences.com/nvidia-gtc-olaf-robotic-character/ ↩
12. TechCrunch, "Nvidia and Google DeepMind will help power Disney's cute robots," March 18, 2025. https://techcrunch.com/2025/03/18/nvidia-and-google-deepmind-will-help-power-disneys-cute-robots/ ↩
13. VentureBeat, "NVIDIA announcements, news and more, from GTC 2025," March 2025. https://venturebeat.com/ai/nvidia-announcements-news-and-more-from-gtc-2025
14. Maginative, "NVIDIA, Google DeepMind, and Disney Research Team Up for Open-Source Physics Engine," March 2025. https://www.maginative.com/article/nvidia-google-deepmind-and-disney-research-team-up-for-open-source-physics-engine/ ↩
15. Dataconomy, "NVIDIA Launches Newton 1.0 Physics Engine For Industrial Robot Training," March 17, 2026. https://dataconomy.com/2026/03/17/nvidia-launches-newton-1-0-physics-engine-for-industrial-robot-training/ ↩
16. GamesBeat, "Nvidia brings graphics research breakthroughs to simulation and physical AI," July 2026. https://gamesbeat.com/nvidia-brings-graphics-research-breakthroughs-to-simulation-and-physical-ai/ ↩
17. Deadline, "Nvidia CEO Chats Up Star Wars-Inspired Droid As Chip Maker Partners With Disney, Google," March 2025. https://deadline.com/2025/03/nvidia-ceo-jensen-huang-ai-disney-next-generation-entertainment-robots-1236329770/ ↩
18. Variety, "Disney Imagineering Moves the AI Bar by Encouraging BDX Droids 'to Emote,'" 2025. https://variety.com/2025/biz/news/disney-imagineering-ai-droids-learning-1236460286/ ↩