Newton v1.6.0 is a feature release following v1.5.1. It improves collision performance and scheduling, expands OpenUSD-authored particle and deformable workflows, adds graph-friendly robot controllers, introduces compliant mechanics and reusable rod modeling, and broadens the capabilities of the experimental Kamino solver.
\nFor the complete list of changes, see the changelog.
\n\nFaster, configurable collision detection. CUDA collision work is faster across broad-phase, narrow-phase, and deterministic contact processing. Solvers can schedule rigid and soft self-contact detection, while opt-in predictive contacts help reduce tunneling between checks. (#4063, #3828, #3854)
\nOpenUSD particles and surface extraction. Newton can import authored MPM particles, materials, and scene configuration from USD and align supported deformable assets with the updated AOUSD proposal. The new ParticleSurface workflow turns particle simulations such as fluids into compact meshes for visualization and export, including multi-world and CUDA-graph use cases. (#3869, #3988, #3349)
Experimental: graph-friendly robot control. New Differential IK and operational-space controllers support heterogeneous robot batches, joint-limit and null-space control, and hybrid force/motion tasks. Implicit actuator response, faster CUDA forward kinematics, and explicit actuator-state copying make more control loops reusable inside CUDA graphs. (#4129, #4105, #3855, #3395, #4101)
\nCompliant robots, rods, and mechanisms. An opt-in compliant ALM mode for SolverVBD covers contacts, structural joints, drives, and limits. The new newton.Rod type provides validated, reusable inputs for ordered chains and explicit graphs, with material properties or section rigidities converted into local joint stiffness. (#3769, #4009)
Experimental: more capable Kamino simulation. SolverKamino adds Coulomb joint friction, joint effort limits, body motion roles, device-resident per-world solve status, and broader batched kinematics support, making multi-world robot simulations easier to configure and diagnose. (#3966, #3990, #3937, #3916)
Faster startup and replicated MuJoCo setup. Cold example-load benchmarks improved by 13% to 20% on tested x86 and ARM systems. Across 8,192-world Ant, Cartpole, G1, and Humanoid workloads, MuJoCo solver initialization was 44% to 77% faster; 256-world G1 and Humanoid setup improved by 12% to 22%. (#3839, #3825, #4119)
\nFaster, lighter model finalization. Shape-BVH construction now uses tiled CUDA reductions for meshes and Gaussian splats; in a measured 4,096-environment rough-terrain workload, this reduced BVH build time from about 903 ms to 17 ms. On ARM, model-initialization CPU memory also fell by 14% for G1 and 25% for Humanoid. (#4001)
\nFaster collision-heavy simulation. Cross-platform benchmarks measured hydroelastic contact simulation 33% to 37% faster, SDF contact simulation 9% to 23% faster, and pyramid contact simulation 11% to 14% faster. CUDA collision detection, mesh preprocessing, convex support queries, heightfield traversal, and deterministic contact processing all perform less work while preserving contact results. (#4063, #3801, #3961)
\nLower overhead and memory at scale. A 64-world hydroelastic workload reduced GPU buffer use from about 54.5 GiB to 3.75 GB while preserving all expected contacts. The tested 8,192-world G1 workloads also spent 30% to 40% less time in Newton-side simulation overhead, and forward kinematics now runs across articulation joints in parallel on CUDA. (#4040, #3395)
\nJoint position-target layout. newton.use_coord_layout_targets now defaults to True, so joint_target_q has one entry per joint coordinate and matches joint_q. Index it with Model.joint_target_q_start; set the flag to False before building models only as a temporary compatibility path. (#4096)
Collision behavior. Mesh SDF construction now filters fully concave manifold edges by default; pass edge_concave_filter=False to retain the previous contact set. Hydroelastic contacts now use margin for interaction distance and gap for detection slack, and XPBD restitution is solved per contact manifold, so re-test scenes tuned around the previous behavior. (#4171, #3719, #4103)
USD deformable and cable assets. The importer adopts the newer AOUSD deformable proposal, including geometry-owned typed thickness and mass arrays, structural stiffness values, and new fallback values. Author array values together with their elementType, move thickness to simulation geometry, and re-test assets with unauthored thickness or uneven cable sampling. (#3988, #3730)
Keyword-only options. Optional arguments across mature builder, geometry, contact, solver, and selection APIs that were deprecated as positional in v1.4.0 must now be passed by keyword. (#4070)
\nMesh adjacency. The deprecated MeshAdjacency.indices, dict-based edges, MeshAdjacency.Edge, and add_edge() APIs are removed. Use tri_indices, edge_indices, and edge_tri_indices. (#4110)
Joint impedance controllers. Deprecated index-port overrides and legacy model-free size arguments are removed. Bind indexed array views directly through the controller's resolved coordinate and DOF indices. (#3979)
\nSolver compatibility APIs. The deprecated Model.mujoco.dof_passive_damping alias and VBD dahl_defaults_enabled parameter are removed. Use Model.joint_damping and explicitly author positive vbd:dahl_eps_max and vbd:dahl_tau values where Dahl rod friction is required. (#3623, #4024)
Rod and cable construction. Migrate from JointType.CABLE, add_joint_cable(), positional add_rod(), add_rod_graph(), CableStiffness, and the cable-construction helpers to JointType.ROD, add_joint_rod(), newton.Rod, and ModelBuilder.add_rod(rod=...). The old forms remain functional through 1.6 and are eligible for removal in 1.7. (#4009)
VBD compliant constraints. Adopt the compliant ALM path with rigid_compliant_alm=True and finite authored stiffness. Pass False explicitly only while migrating from the deprecated legacy hard/soft constraint controls; the default will change to True. (#3769)
Collision ranges and scheduling. Migrate legacy VBD contact-range and detection-interval names to soft_contact_gap, particle_self_contact_margin plus particle_self_contact_gap, and solver-owned collision schedules. (#3828)
Experimental actuator terminology. Migrate the experimental actuator Controller* classes and controller* fields to Drive* and drive*, use ClampingBase, and use ComponentKind.DRIVE. The former names continue to warn during the 1.6 deprecation window. (#4054)
See the changelog for the complete set of new deprecations, including USD deformable attributes and keyword-only migrations for Mesh, TetMesh, and newton.usd.get_mesh().
Newton now requires warp-lang>=1.17.0.
The sim extra moves to MuJoCo and MuJoCo Warp 3.12.
USD import workflows require newton-usd-schemas>=0.5.0.
The onnx extra moves to warp-nn[onnx]==0.3.1, and documentation and example workflows require GitPython 3.1.58 or newer.
No new required dependency names were added. For complete dependency details, see the pyproject.toml comparison.
Thanks to the Newton Project Members and to the following contributors from outside the Newton maintainer, TSC, and project-member groups:
\n@han-xudong for MJCF actuator and tendon import improvements, Kamino terminal status, triangle-metadata correctness, and the asRoBallet control example. (#3669, #3684, #3685, #3744, #3746, #3937, #3956)
\n@ooctipus for predictive rigid contacts, OpenUSD texture-transform preservation, and corrected floating-base IK Jacobians. (#3854, #4007, #4088)
\n@ruziniuuuuu for USD collider-approximation handling, mesh-normal input support, and avoiding unused initialization contacts in MuJoCo. (#3831, #3833, #3825)
\n@deepspace28 for restoring MuJoCo line-search performance on large models and keeping VBD proxy reaction forces within configured contact capacity. (#3908, #4004)
\n@rakhimovv for bounded headless ViewerGL execution, headless-rendering guidance, and VBD edge/face contact validation. (#3930, #3932, #4060)
\n@aneangel for Gaussian-splat rendering and stable plane grids in ViewerViser, plus clearer MuJoCo free-joint drive warnings. (#4018, #4019, #4083)
\n@LuckyIYI for reducing unnecessary broad-phase work around infinite planes. (#3867)
\n@CharlesKZW and @Double7sBurger for MJCF heightfield orientation and heightfield-contact correctness fixes. (#3898, #4005)
\n@CoffeeDrivenCoder for supporting non-int32 mesh indices in normal computation, @sparkle2820 for restoring Python 3.10 ModelBuilder construction, and @zlong-911 for NumPy 1.x compatibility in anisotropic Style3D cloth. (#3986, #3942, #4027)
@liorbenhorin for OpenCV pinhole camera-ray generation and @mcrowleyNV for the mobile documentation navigation fix. (#3901, #4082)
\n