Abstract
The Kamino simulation software is a GPU-accelerated physics solver designed for accurate simulation of complex mechanical systems with kinematic loops, such as parallel manipulators, planar and spatial linkages, and multi-limbed robots with coupled joints, without requiring simplified approximations. Built on NVIDIA Warp and integrated into the open-source Newton physics framework, the Kamino simulator enables large-scale reinforcement learning by running thousands of parallel environments on a single GPU, including heterogeneous worlds where each environment can contain a structurally different robot.
Overview
Simulating complex mechanical systems, such as multi-limb robots, is essential for development and testing, especially when using deep reinforcement learning (RL). Practitioners often face difficulties with highly coupled mechanical assemblies that contain intrinsic closed kinematic loops, which are frequently used to transfer power more efficiently than serial chains. Simulating these systems involves challenges such as ill-conditioning due to large mass ratios, redundant constraints, and the need for accurate constraint satisfaction and stabilization. Consequently, these kinematic complexities are often approximated by equivalent articulated systems, which requires expertise and can increase the sim-to-real gap.
Figure: Most physics simulators focus on kinematic trees (open chains) and sometimes provide a limited set of loop closure constraints. The Kamino simulator natively supports arbitrary joint topologies, including closed kinematic loops, enabling accurate simulation of complex mechanical assemblies.
When to Use the Kamino Simulator
Note: The Kamino simulator is currently in beta. Performance improvements and support for additional features are actively being worked on.
The Kamino simulator is designed for simulating complex mechanical systems that go beyond simple kinematic trees. Consider using the Kamino simulator when your application involves:
- Kinematic loops — closed-chain mechanisms such as planar and spatial linkages, parallel manipulators, and multi-limbed robots with coupled joints. The Kamino simulator simulates the actual mechanical assembly rather than requiring approximations.
- Accurate and robust dynamics — The Kamino simulator is based on an ADMM (Alternating Direction Method of Multipliers) solver that provides accurate constraint satisfaction and handles ill-conditioned systems with large mass ratios, redundant constraints, or tightly coupled interactions between bodies.
- GPU-accelerated batch simulation — training reinforcement learning policies by simulating many parallel worlds on a GPU. Unlike most simulators, the Kamino simulator supports heterogeneous environments where each parallel world can contain a completely different model (different bodies, joints, and collision geometries).
When Not to Use the Kamino Simulator
- Pure articulated systems — if your system is a kinematic tree without closed loops, a dedicated articulated-body solver will be faster. The Kamino simulator's generality comes at a computational cost.
- Low-latency, single-system simulation — simulating a single environment in real time is not the Kamino simulator's strength. For low-latency simulation, a dedicated CPU simulator is preferred.
Ecosystem
The Kamino simulator is a solver backend within NVIDIA Newton, an open-source framework for GPU-accelerated physics simulation. Newton provides a common interface for multiple solver backends, and the Kamino simulator extends this with native support for constrained multi-body systems with kinematic loops.
For reinforcement learning workflows, integration with Isaac Lab, NVIDIA's framework for robot learning, is planned and will be available soon. Isaac Lab will provide environment wrappers, task definitions, and training pipelines on top of Newton's physics backends, enabling end-to-end RL training with the Kamino simulator.
Basic Examples
Robot Examples
Basic Usage
Loading a model from USD
The simplest way to get started is to load a model from a USD file:
Advanced Setup
The Kamino simulator examples showcase some of the features beyond a basic setup:
- All examples feature a multi-world setup with CUDA graph capture to speed up GPU-based computations.
- The fourbar example shows how to use the simulation configuration to adapt simulation parameters.
- The DR Legs example sets up PID control to actively move the robot based on an input animation.
Acknowledgements
We thank Josefine Klintberg, Espen Knoop (Disney Research, Zurich) and Kyle Laughlin (Walt Disney Imagineering R&D, Glendale), as well as Gilles Daviet, Eric Heiden, Miles Macklin, Gavriel State, Alain Denzler, Philipp Reist, Adam Moravanszky, and Spencer Huang (NVIDIA) for their contributions and support.
Citing the Kamino Simulator
If you use the Kamino simulator in your research, please cite the following papers:
Kamino simulator system description: coming soon.
@article{kamino2026,
title={Kamino: GPU-based Massively Parallel Simulation of Multi-Body Systems with Challenging Topologies},
author={Vassilios Tsounis, Guirec Maloisel, Christian Schumacher, Ruben Grandia, Agon Serifi, David M{\"u}ller, Chris Amevor, Tobias Widmer, Moritz B{\"a}cher},
year={2026},
}
Algorithmic foundations: the core solver algorithm and its mathematical formulation.
@article{tsounis2025solving,
title={On Solving the Dynamics of Constrained Rigid Multi-Body Systems with Kinematic Loops},
author={Tsounis, Vassilios and Grandia, Ruben and B{\"a}cher, Moritz},
journal={arXiv preprint arXiv:2504.19771},
year={2025},
}