Field noteNov 16, 20265 min read1 source

Newton 1.8.0 ships experimental SolverSteel from Skild AI

In one minute

Newton 1.8.0 adds the experimental SolverSteel to its physics engine, alongside new tendon, deformable-simulation and rendering capabilities. Developed with Skild AI, Steel targets high-throughput, memory-efficient simulation for contact-rich reinforcement learning; its experimental status means teams should validate it against their own scenes before relying on it.

What happened

The Newton project’s v1.8.0 release, published on November 11, describes the update as a feature release after v1.7.0. Its headline addition is SolverSteel, an experimental solver developed in collaboration with Skild AI. The release says Steel targets high-throughput, memory-efficient simulation for contact-rich reinforcement learning. It handles environments with multiple rigid bodies and articulations through a matrix-free projected Gauss-Seidel solver. The notes also name run-to-run determinism, island-based sleeping and patch-based Coulomb friction as supported features. The release does not give a performance figure for Steel itself.

The update also introduces experimental massless routed tendons. Newton says these can model compliant routes through anchors, pinholes and rollers, including capstan friction and dynamic routing, in XPBD and VBD without representing a massive cable as a discretized chain. For deformables, v1.8 adds import of authored cable, cloth and volume rest shapes independently of their starting poses. Mesh feature contacts and an extended VBD rigid-soft Divide-and-Truncate method are intended to improve contact between coarse deformables and detailed rigid meshes.

On the visual side, experimental skinned deformable visuals let graphics meshes attach to simulated cables, cloth or volumes without raising physics resolution. The release describes importing USD bind poses and textures and sharing evaluated geometry across supported viewers and camera RGB/depth output; interactive skinned meshes use GL or RTX. Other changes include filterable Newton warnings and deprecations, library diagnostics through Python logging, heterogeneous articulation layouts across worlds, and a surface-gripper example runnable with python -m newton.examples surface_gripper.

The release calls out migration work as well as features. Among the changes, XPBD stiffness behavior changes with multiple solver iterations; URDF import now defaults to passive damping for XPBD and VBD; and VBD defaults to compliant ALM for rigid contacts and joints. CUDA fast math is enabled by default. Python 3.10 remains supported in 1.8 but is deprecated for removal in 1.9. Newton also documents breaking API changes, solver option migrations and dependency constraints, including MuJoCo 3.17 for the simulation extra and Warp 1.19.0 for release development and validation; the package minimum remains Warp 1.17.0.

What the archive adds

New Runtime covered Standard Physics’ Archimedes design agents and Reynolds on November 6, 2026, describing an engineering design-agent system and an open-source CFD agent. That is related through engineering simulation, but Newton 1.8.0 is a new event: a physics-engine release rather than another announcement from Standard Physics.

Trend signal

This release supports scientific agents need artifact evidence only weakly: reproducible, deterministic simulation can matter to evidence generation, but the release does not describe an agent or an evidence package. More directly, this does not yet fit a pattern New Runtime tracks; it is a specialized simulation-engine release, not enough by itself to establish a broader shift.

Why now

The release is the first feature update after Newton 1.7.0 and bundles the Steel solver with migration changes that affect existing projects. Teams evaluating it need to account for both the new experimental path and changes to established solver behavior.

Why it matters

For teams building physical agents, contact simulation can shape how quickly and reliably they generate training experience. SolverSteel’s stated focus on throughput, memory efficiency and articulated, contact-rich scenes makes it relevant to robotics workloads that combine many bodies and interactions. Determinism is also a useful property when debugging a simulation or comparing runs, although the release gives no quantified reproducibility or speed results for Steel.

The broader 1.8.0 changes may matter just as much to adoption. Better deformable contacts, authored rest shapes and visual meshes that do not require higher physics resolution address distinct simulation and presentation needs. But changing defaults and APIs create upgrade risk: teams should review the migration guidance, especially if they rely on XPBD/VBD stiffness, imported damping, custom Rod layouts or prior fast-math behavior. An experimental solver should be benchmarked and validated against representative scenes before it becomes a training dependency.

What happens next

  • Newton will keep SolverSteel experimental until its documentation or release notes describe a stable support contract. (6 months, 65%). Watch for: a later release changing its experimental designation or documenting compatibility guarantees.

Limits and uncertainty

Newton’s release notes do not quantify SolverSteel’s throughput, memory use, scaling, or accuracy relative to Newton’s other solvers or external alternatives. They also do not specify a stability timeline or production-support commitment for this experimental feature. The claimed determinism is not accompanied by a test protocol or scope. Teams should treat the listed capabilities as release claims, then measure behavior on their own scenes, hardware and training workloads. Several other features in 1.8.0 are also explicitly experimental, and platform or dependency combinations can constrain what is usable.

Skill to build now

Build a small, repeatable validation scene before migrating a robotics workload. Include the articulated bodies and contact patterns that dominate the real task, run it repeatedly to check whether outcomes are deterministic, and record runtime and memory use alongside contact behavior. Compare SolverSteel with the solver currently in use, then separately test upgrade-sensitive defaults such as fast math and VBD iteration settings. Keep the scene and results under version control so later Newton releases can be compared on the same basis.

Reader takeaway

Newton 1.8.0 adds a promising but experimental solver for contact-rich articulated simulation, plus meaningful deformable and tendon updates. The practical move is to test it against a representative workload—not to assume the stated targets translate into gains for every scene—and to treat the migration notes as part of the upgrade.

Evidence / sources

  1. 1github.com/newton-physics/newton/releases/tag/v1.8.0
All Field Notes

Next reads

Read next

Open the graph
  1. 01archiveField NotesOpen the latest editorial analysis.
  2. 02sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract