The NDEM engine.
A GPU-native, real-time granular DEM engine — accurate enough that policies transfer to real ground, fast enough to train them at scale.
Discrete element modelling captures granular physics faithfully — but engineering-grade DEM is slow. For iron-fines simulation, a typical timestep is around 1×10⁻⁶ s, which means hours of compute for every second of simulated digging. At that speed, closed-loop policy training is simply out of reach.
NDEM runs stably at timesteps around 1×10⁻⁴ s — roughly 100× larger — without blowing up, with granular behaviour that stays true from free-flowing grain to sticky, cohesive ore. That is the difference between a research curiosity and a training engine.
Designed for RL workflows from day one.
Not a solver bolted onto a training loop after the fact — large-scale parallel throughput is what the architecture is built around.
Built for parallel RL
Headless batch simulation, designed to slot into large-scale reinforcement-learning training loops.
Domain randomisation
Designed so material parameters — friction, cohesion, size distribution — can be randomised across environments for robust sim-to-real transfer.
Massively parallel
Architected to scale to thousands of granular environments in parallel, faster than real time.
Deterministic replay
Designed for exact rollout reproduction — essential for debugging policies and validating results.
A first-class granular solver for NVIDIA robotics.
NDEM is built on Warp and couples with Newton — the native path into Isaac Lab and Isaac Sim, where robotics teams already train.
GPU-native kernels. NDEM is written directly on Warp, so it runs where the robot-learning workloads already live.
We have demonstrated two-way coupling with Newton’s Featherstone solver — a robot arm digging in granular media, forces flowing both ways.
The native path to upstream NDEM as a first-class granular solver in NVIDIA’s robot-learning ecosystem.
The data most robotics teams can't get.
Deep material-calibration experience
Two decades of DEM and bulk-materials fieldwork calibrating granular materials against measured site data — the kind of experience most robotics teams cannot access. We are looking to build that into a validated-material database together with partners and clients; see the roadmap below.
Digital twins from CAD
Model the actual terrain, muck pile, or bucket geometry from CAD — the real operation the robot has to work in, not a simplified stand-in.
NDEM demonstrating two-way coupling with Newton’s Featherstone solver — a robot digging in granular media.
Engine benchmark release, and the NDEM solver in Newton + Isaac Lab (beta).
First robotics-customer pilot, building the calibrated-material database with partners and clients, site by site.
Bringing granular simulation into your robotics stack?
RL training environments, digital twins, or a Newton / Isaac Sim integration — tell us what you're modelling and we'll help you get NDEM into your pipeline.