Newton v1.5.0 adds vectorized joint control for robot simulation
Newton v1.5.0 has been released on the project’s public GitHub repository, adding experimental vectorized joint control, broader batched simulation support and a long list of contact, cable and asset import fixes for robot learning workflows.
The release is relevant for teams using simulation to train or validate articulated robots at scale. According to the project’s GitHub release notes, the new newton.controllers module adds model based and caller supplied joint space impedance controllers for heterogeneous GPU robot batches, with a runnable example included.
Batched robot workloads get attention
Several changes target multi world simulation, a common pattern in reinforcement learning and policy testing. Newton v1.5.0 adds isolated implicit MPM, masked solver resets and dedicated global world gravity. The project says representative replicated Humanoid and G1 workloads initialize 11% to 24% faster, while CPU peak initialization memory falls 10% to 15% across Cartpole, Humanoid and G1.
The release also adds opt in MuJoCo Warp sleeping. Tree sleeping, contact based wake up, initial sleep policies and compact nvmax storage are intended to reduce active work and solver memory in scenes where many articulated trees can remain inactive.
More contact and import fixes
Contact behavior is another major theme. Newton v1.5.0 includes deterministic hydroelastic generation, persistent geometry, selective reset, stable box manifolds, full surface VBD proxy coupling and Kamino capacity fixes. For robotics users, these are less glamorous than a new model or benchmark, but they are the kind of changes that can decide whether a large batch run is reproducible or fragile.
Cable mechanics also get a substantial update. The release adds helpers and experimental VBD slots that model stretch, shear, bend and twist independently, along with seven validation examples and a plectoneme showcase.
Asset import fidelity improves across USD and MJCF. The release notes say importers now preserve purpose, visibility, materials, textures, mass, velocity, physics scene discovery and selection, collision filtering, sites and equality constraints more faithfully. Faster terrain and USD construction are also listed as setup improvements.
Migration work is required
This is not a drop in update for every codebase. Newton v1.5.0 removes the deprecated joint_target_pos and joint_target_vel aliases from Model, Control and ModelBuilder. Users are directed to joint_target_q and joint_target_qd.
The release also deprecates scalar gravity, model owned collision helpers and DOF shaped position targets for models with free, ball or distance joints. Newton now requires warp-lang>=1.16.0, the sim extra moves to MuJoCo and MuJoCo Warp 3.11, and USD import workflows require newton-usd-schemas>=0.4.1.
Source: share.google
