Newton v1.4.0

Newton v1.4.0 is a feature release centered on composing solvers, making captured and deterministic workflows more portable, expanding articulated dynamics tooling, improving deformable simulation and USD import, and delivering targeted performance gains in graph-captured and large-scale workloads.

For the complete list of changes, see the [changelog](https://github.com/newton-physics/newton/blob/v1.4.0/CHANGELOG.md#140.

Highlights

• Experimental coupled solver composition. SolverCoupled, proxy coupling, and ADMM contacts make it possible to compose MuJoCo, Kamino, VBD, XPBD, and MPM within a single workflow. (#2848)

• Experimental deformable USD workflows. Based on our AOUSD schema proposals, ModelBuilder.add_usd() can import curves, meshes, tetrahedral meshes, attachments, and element filters as Newton cables, cloth, soft bodies, joints, and collision filters. Tiled-camera rendering now treats cloth and volume topology as meshes without also rendering their vertices as particle spheres. (#3192, [#3518])

• Deterministic and portable execution. XPBD, SemiImplicit, Featherstone, VBD, and MuJoCo can opt into deterministic kernel paths for bit-exact repeated rollouts. Supported examples can also capture and replay Warp graphs on CPU. Default solver execution remains unchanged. (#3300, [#3387])

• VBD deformable simulation improvements. VBD gains opt-in full-surface rigid-soft contact that detects soft-triangle edge and face crossings against rigid SDFs, closing gaps that per-particle detection can miss. It also adds per-shape soft-contact margins and more consistent damping behavior across deformable materials, springs, cables, joints, and contacts. (#3262, [#2994], [#2877])

• Articulated inverse dynamics. The experimental eval_inverse_dynamics_passive() and eval_inverse_dynamics_force() APIs populate caller-owned arrays with joint-space mass matrices, gravity and Coriolis forces, and manipulator-equation joint forces for full models or selected articulations. (#2753, [#3530])

Performance improvements

• Much faster launch-heavy CPU workflows. Warp 1.15 CPU graph capture can greatly reduce launch overhead for graphable Newton workloads. A representative XPBD quadruped workload runs about 16 times faster with graph replay. (#3387)

• Higher-throughput batched MuJoCo simulation. Large GPU workloads for G1, KitchenG1, and Humanoid run about 15% to 26% faster, while challenging Allegro workloads improve by about 11% with MuJoCo Warp 3.10.0.2. (#2980, [#3502])

• Faster simulation and visualization paths. Representative cloth and contact-heavy scenes run about 10% to 13% faster, and ViewerGL rendering is around 10% faster for large multi-world scenes.

Announcements

Breaking changes and removals

• MuJoCo gap propagation and import semantics changed. SolverMuJoCo now propagates authored shape_gap into MuJoCo geom_gap and pair_gap, while MJCF and USD imports no longer subtract gap from margin by default. Shapes without an explicit gap inherit ModelBuilder.rigid_gap = 0.1 m, which now reaches MuJoCo and may increase detection-only contacts, broad-phase work, and contact-buffer pressure. Use legacy_margin_gap=True for assets authored against MuJoCo 3.8 or earlier. For other scenes, retune builder.rigid_gap before adding shapes or set ShapeConfig.gap per shape; 0.0 preserves the previous MuJoCo detection envelope. (#2980)

• USD material resolution now follows canonical bindings only. Newton no longer falls back to non-canonical material relationships. Assets that rely only on bare purpose-specific bindings such as material:binding:preview, or bind to prims that are not Material prims, may stop resolving materials. Validate and repair affected assets with usdchecker or usd-validation-nvidia. (#3350)

• Several simulation defaults and semantics changed. The CoACD threshold now defaults to 0.05; standalone SolverSemiImplicit enables particle-particle contact forces; SolverFeatherstone solves floating bases about the root center of mass; and inactive MPM particles no longer participate in grid transfers. Review the changelog for opt-outs and migration guidance before reusing tuned scenes.

• Implicit MPM collider setup is keyword-only. All arguments to SolverImplicitMPM.setup_collider() must now be passed by name. Migrate positional calls to keywords such as collider_meshes=, collider_body_ids=, and collider_particle_ids=. (#3524)

• APIs deprecated in earlier releases were removed. Notable removals include SensorRaycast, the SensorTiledCamera 1.1 compatibility surface, the external newton-actuators dependency and legacy actuator signature, top-level equality-constraint compatibility fields, and body armature. Migration paths are listed in the changelog. (#3169, [#3168], [#3170], [#3296], [#3153])

New deprecations and upcoming removals

• Option-heavy solver, builder, shape, joint, and FEM APIs now warn when options are passed positionally. Keep stable identifiers positional and migrate configuration arguments to explicit keywords. (#2993)

• Migrate from SensorTiledCamera.config and render_config to default_render_config, set texture and shadow behavior explicitly, and use the renamed camera-ray helpers. Other newly deprecated surfaces include State.body_q_prev, legacy neural-network checkpoint formats, implicit rod body-frame selection, and the legacy contact-match constants. See the changelog for the complete migration list. (#3408, [#3372], [#3356], [#3498])

Dependency updates

• Newton now requires warp-lang>=1.15.0.
• The simulation extra moves to MuJoCo 3.10 and mujoco-warp>=3.10.0.2,<3.11.
• USD import workflows require newton-usd-schemas>=0.4.0.
• The new onnx extra adds ONNX policy inference through warp-nn[onnx]==0.3.0.
• The deprecated newton-actuators dependency is removed; use newton.actuators.

For complete dependency details, see the pyproject.toml comparison (github.com).

Acknowledgments

Thanks to the Newton Project Members and to the following contributors from outside the Newton maintainer, TSC, and project-member groups:

• @han-xudong for URDF package resolution, viewer-state, SemiImplicit, XPBD-contact, and restitution fixes. (#2614, [#2658], [#3204], [#3304], [#3287])
• @aneangel for plotting and ViewerUSD improvements. (#2672, [#2990], [#2992])
• @devshahofficial for the example render-rate cap and USD mesh loading. (#2606, [#3214])
• @maxkra15 for clearer USD schema failures and batched MuJoCo geometry fixes. (#3117, [#3376])
• @felixmey for stabilizing and regression-testing the cable cross-slide table example. (#3126)
• @huidongc for fixing triangle-mesh construction in RenderContext. (#3199)
• @Langford-H for fixing ViewerFile running state after close. (#3190)
• @haoran-whynot for fixing Style3D memory growth without CUDA Graph capture. (#3243)
• @rshahid-nvidia for exposing BVH constructor selection in ModelBuilder. (#2864)
• @Dev-Jahn for fixing Implicit MPM graph capture with iterative linear solvers. (#3191)
• @jpatel3 for preserving unrelated USD articulations during import. (#3218)
• @mehdiataei for fixing CPU frame capture in ViewerGL. (#3093)
• @hujc7 for honoring ignored paths during custom-frequency USD traversal. (#3406)
• @ooctipus for canonical USD material-binding resolution. (#3350)