Introduction
A number of announcements from NVIDIA's keynote and technical papers at SIGGRAPH 2026 (July 19–23, Los Angeles) caught my attention from a practical standpoint. Among them, the new solver for the Newton physics engine stood out to me, since I regularly work with Isaac Sim and Isaac Lab.
I had just finished a pick-and-place simulation using the OnRobot 2FG7 gripper, so I wanted to take this opportunity to think through "what would have changed if this had been available six months ago."
What Is Newton?
Newton is an open-source, GPU-accelerated physics simulation engine developed jointly by NVIDIA, Google DeepMind, and Disney Research, and governed under the Linux Foundation. It is already integrated into the develop branch of Isaac Lab 3.0 Beta, where it can be tried out as a backend for reinforcement-learning- and imitation-learning-based robot policy training.
To be precise, it isn't accurate to say "Newton was announced at SIGGRAPH 2026." What was actually announced at SIGGRAPH 2026 — introduced as one of NVIDIA's 21 accepted technical papers — was a new solver capable of handling hard-to-simulate materials such as snow, sand, and elastic solids in real time, which was released as an addition to the Newton engine (NVIDIA Blog). The Newton engine itself has been in development and publicly available for some time before this.
The biggest difference from the physics backends Isaac Sim/Isaac Lab has used until now (PhysX 5 / Warp) is said to be the accuracy of contact dynamics and the efficiency of GPU-parallel processing. In pick-and-place tasks — grasping and placing an object — how the contact surface between a gripper and the target object is physically modeled has a major impact on simulation accuracy. Newton appears to be designed to compute this on the GPU with both high precision and high speed.
The Relationship with the Sim-to-Real Gap
Anyone who has worked with robots in Isaac Sim inevitably runs into the sim-to-real gap: the object grasps perfectly in simulation, but on the real robot it slips, gets dropped, or ends up in the wrong position.
There are several possible causes, but from a physics standpoint the following two are common culprits.
- Simplified contact models: contact in simulation is physically simplified and doesn't reproduce real-world rubber, friction, or deformation
- Physical consistency across parallel environments: when training across a large number of parallel environments, variation in physics behavior between environments can hurt policy generalization
Newton is said to address both of these. Improvements to contact stiffness and friction-coefficient modeling, combined with GPU-parallel processing that keeps physical consistency high across environments, could plausibly help close the gap.
What the "Real Gap" Looked Like in My Own PoC
A while back, I built a pick-and-place environment on Isaac Sim 5.1 and ran a proof of concept connecting it in real time via Modbus TCP (details here).
The part of that sim-to-real work that took the most effort was the gripper's grasp timing and contact response. Even though the object appeared stable the moment it was grasped in Isaac Sim, there was a mismatch between the timing of the DO signal sent back to the controller over Modbus and the actual moment the grasp was complete. I worked around it by adjusting the polling interval, but at its root this is also a question of how accurately the physics simulation detects contact in the first place.
If Newton improves contact-model accuracy, it could make it easier to deal with this kind of question: how do you define "grasp complete" within the simulation?
Other Notable Announcements
Here's a quick summary of a few other announcements I found practically relevant beyond Newton.
Cosmos 3 Edge A 4-billion-parameter open-source world foundation model. It takes text, images, video, audio, and actions as input, reasons about the physical scene, and outputs robot actions. It runs across a wide range of hardware — Jetson, RTX PRO, DGX, and GeForce RTX — and NVIDIA has published the weights, inference code, and post-training recipes on Hugging Face (details).
Note that the demo "running on a single RTX PRO 6000" shown during the keynote was not Cosmos 3 Edge itself, but Cosmos-Dreams, an autonomous-vehicle simulator. The point of that demo was compressing a workload that previously required 64 GB300 GPUs down to a single RTX PRO 6000 — a separate announcement from Cosmos 3 Edge (NVIDIA Blog).
Is Newton "Ready to Use" Right Now?
According to the official documentation, a locomotion policy trained with Newton has already been deployed to a real Unitree G1 robot, and policy transfer between Newton and PhysX has also been validated. In other words, some sim-to-real validation has already taken place. However, this is mainly for locomotion (walking) tasks — the contact-dynamics improvements relevant to grasping tasks like the pick-and-place work discussed in this article don't appear to have been validated yet.
For teams currently building pick-and-place learning environments in Isaac Lab, switching an existing environment over to Newton once it lands in a stable release could improve how well trained policies transfer to real hardware.
Summary
- Newton is an open-source, GPU-based physics engine jointly developed by NVIDIA, Google DeepMind, and Disney Research. It's already integrated into Isaac Lab 3.0 Beta; what was announced at SIGGRAPH 2026 was a new solver for snow, sand, and elastic materials
- For locomotion tasks, deploying a Newton-trained policy to a real G1 robot has already been validated. Contact-dynamics improvements for grasping tasks, however, are not yet validated
- The timeline for stable inclusion is still undetermined, but it's worth factoring the design direction into planning now
- Cosmos 3 Edge could make edge inference more practically viable. The single-RTX-PRO-6000 demo belongs to the separately announced Cosmos-Dreams, so the two shouldn't be conflated