Brain / model / chipBrain

Newton 1.5 physics engine

NVIDIA (United States) ·
Illustration

No official image on file. Showing a schematic category illustration.

Release of the Newton physics engine with more parallel simulation, lower memory and consistent contact physics for robot training.

Key specs

SpecificationsNot stated

Technical details

Newton v1.5.0 (released 2026-08-11) is a feature release of the open-source Newton physics engine aimed at scalable batched simulation for robot learning. It adds an experimental newton.controllers module with vectorized joint-space impedance control for heterogeneous GPU robot batches, isolated implicit MPM, masked solver resets and per-world gravity for multi-world workflows, and opt-in MuJoCo Warp sleeping. Contact handling is more reproducible (deterministic hydroelastic generation, stable box manifolds), cables gain independent stretch/shear/bend/twist, and USD/MJCF import fidelity improves. Code is Apache-2.0.

ArchitectureGPU-accelerated, differentiable-oriented physics engine for robot learning with multiple solver backends (MuJoCo Warp, VBD, implicit MPM, Kamino) and USD/MJCF importers.
Inputs outputsInputs: robot/scene assets (USD, MJCF) and control commands; outputs: batched simulation state for multi-world robot learning.
Embodiments supportedArticulated robots imported via USD/MJCF (general-purpose).
License codeApache-2.0
Hardware requirementsNVIDIA GPU with CUDA (CUDA graph capture supported).
How to trygithub.com/…/newton
Sources

Benchmarks, where listed, are as claimed by the vendor.

Sources

More in Brain / model / chip

Illustration
Pi API (beta)

Pi API (beta)

Physical Intelligence · Brain / model / chip
NVIDIA Isaac ROS 5.0Open source

NVIDIA Isaac ROS 5.0

NVIDIA · Brain / model / chip
NVIDIA Jetson Orin Nano 2

NVIDIA Jetson Orin Nano 2

NVIDIA · Brain / model / chip