NVIDIA, Google DeepMind, and Disney Built a Physics Engine to Train Every Robot on Earth. Here Is What Newton Does.

Three organisations that do not typically build things together — NVIDIA, Google DeepMind, and Disney Research — jointly developed an open-source physics engine called Newton and donated it to the Linux Foundation. Jensen Huang announced the collaboration at GTC 2025. In 2026, the implications of what they built are becoming clearer.

Newton is not a consumer product. It is infrastructure for training robots — specifically, the simulation environment where robots learn how to interact with the physical world before they are deployed in it. Understanding why this matters requires understanding the single hardest problem in modern robotics.

The problem Newton is solving

Teaching a robot to do something physical — pick up an object, walk on uneven terrain, perform a handshake, do a somersault — requires millions of attempts. In the real world, millions of attempts means millions of hours and millions of potential robot failures. That is why robotics research has moved to simulation: you train the robot inside a computer, let it fail millions of times at essentially zero cost, and then transfer the learned behaviour to the real robot.

This approach works. It is how Unitree's humanoid robots learned to do autonomous kung fu at China's Spring Festival Gala. It is how Boston Dynamics trains Atlas to run and jump. It is how industrial robots learn manipulation tasks that would take human trainers months to teach through demonstration.

The problem is simulation fidelity. The physics in a simulator is never exactly the physics of reality. Friction behaves differently. Objects deform in ways the simulation did not model. Motors have backlash and imprecision that the simulated motor does not. When you take a robot trained in simulation and put it in the real world, it often fails — not because the learning was wrong, but because the simulation was not accurate enough.

This gap between simulation and reality is called the sim-to-real transfer problem, and it is the central unsolved problem in practical robotics deployment.

What Newton actually does

Newton is an open-source physics engine built specifically to close the sim-to-real gap. It was co-designed by three organisations that each brought something critical:

NVIDIA contributed Warp — a GPU-accelerated Python framework for physics simulation that runs on NVIDIA hardware. Newton is built on top of Warp, which means simulations run at GPU speed rather than CPU speed. This is not a marginal improvement. Physics simulations that previously took hours can complete in minutes.

Google DeepMind contributed MuJoCo-Warp — a GPU-accelerated version of MuJoCo, the physics simulator that has been the industry standard for robotics research for over a decade. The collaboration between Newton and MuJoCo-Warp produces a 70x speedup in robotics machine learning workloads compared to the existing GPU-accelerated simulator. When you can run 70x more training iterations in the same time, the quality of what the robot learns improves dramatically.

Disney Research contributed their experience training robots that need to operate in the most demanding physical environment imaginable: a theme park, surrounded by thousands of humans, in real time, with the need to be entertaining and safe simultaneously.

The result is a physics engine with differentiable physics (gradients flow through the simulation, enabling more sophisticated learning algorithms), high extensibility for multiphysics simulations (soft bodies, liquids, cables, deformable materials), and compatibility with the existing MuJoCo ecosystem that most robotics researchers already use.

Why Disney's involvement is meaningful

Disney Research is not a charity contributor to this project. They are motivated by a specific problem they need to solve: the BDX droids.

The BDX droids are Disney's next-generation park robots — small, wheeled humanoid robots inspired by the Star Wars universe, specifically designed to be emotionally engaging to park visitors. They are not remote-controlled. They need to navigate unpredictable environments, respond to human interaction, avoid collisions, and do all of this while appearing to have personality and intent.

Training a robot to behave that way in a Disney theme park — with crowds, children, uneven surfaces, unpredictable human behaviour — requires a simulation environment that accurately models all of those variables. Disney Research is using Newton to build that simulation environment.

The broader point is that Disney's use case is arguably the hardest consumer robotics application: a robot that must be safe, unpredictable-environment capable, and emotionally engaging simultaneously. If Newton works for that, it works for most commercial robotics applications.

What the Linux Foundation donation means for developers

Donating Newton to the Linux Foundation means it is now a community project, not an NVIDIA product. The source code is publicly available. Any robotics developer, researcher, or company can use it, contribute to it, and build on it without licensing fees.

This is significant because the tooling cost for robotics development has historically been high. Simulation environments, physics engines, and ML frameworks for robotics were either expensive commercial products or academic tools with limited production readiness. Newton is a frontier-quality physics engine, co-developed by three of the best AI research organisations in the world, available for free.

For developers interested in robotics, this is the equivalent of when PyTorch and TensorFlow became open source — it removes a significant barrier to entry and shifts the competitive advantage from "having access to good tools" to "knowing how to use good tools effectively."

The connection to GTC 2026

NVIDIA's GTC conference runs March 16–19, 2026. Newton was announced at GTC 2025. GTC 2026 will almost certainly include updates to the Newton ecosystem — new capabilities, updated integration with Isaac GR00T N1 (NVIDIA's open humanoid robot foundation model), and announcements about which robotics companies are using Newton in production.

Isaac GR00T N1, announced alongside Newton, is a foundation model for humanoid robots — a pretrained model that robots can fine-tune for specific tasks rather than training from scratch. The combination of Newton (the simulation environment) and GR00T N1 (the pretrained model) is essentially a development stack for building capable humanoid robots, available as open-source infrastructure.

Watch the March 16 keynote for updates. NVIDIA's physical AI track at GTC is specifically about this stack, and Jensen Huang has said the conference will include hardware surprises that change the physics of what is possible for AI developers.

Why three rivals built something together

NVIDIA, Google DeepMind, and Disney are not natural collaborators. NVIDIA competes with Google in AI hardware. Google DeepMind competes with OpenAI (which has a relationship with NVIDIA) in AI models. Disney is a media company with no obvious alignment with either.

The collaboration exists because the sim-to-real problem is so hard that no single organisation has solved it, and solving it requires pooling expertise in GPU compute (NVIDIA), simulation physics and ML (Google DeepMind), and real-world robotic deployment in demanding environments (Disney Research).

The fact that three organisations with competing interests found it worth collaborating on Newton is itself a signal about how central the sim-to-real problem is to the future of robotics. It is not a solved problem that one company keeps proprietary. It is an unsolved problem that benefits everyone if solved openly.

What to watch in 2026

The Newton ecosystem is developing quickly. By the end of 2026, watch for:

• Adoption metrics: how many robotics companies and research labs are using Newton vs. other simulators
• GR00T N1 updates at GTC 2026: what new capabilities are announced for the humanoid robot foundation model
• Disney's BDX droid deployment: whether the droids appear in parks and how they perform in the real world
• Community contributions to Newton: what the open-source community adds to the base engine

The robotics industry is at an inflection point. Simulation quality has been the limiting factor in deploying capable robots at scale. Newton, if it lives up to what its creators demonstrated at GTC 2025, removes a significant part of that constraint. The consequences of that will show up in the kinds of robots that are deployable in 2027 and 2028.