KangEngine

Character animation and GPU-parallel physics simulation in C++ and Python.
Built for robot learning, motion workflows, and interactive visualization.

Task	Framework	Environments	Environment Step FPS	Step + Inference FPS	Step + Inference + Train FPS
Cartpole	KELab (KangEngine)	4096	1,153,317	914,046	527,070
Cartpole	IsaacLab v2.3.2 (local)	4096	978,112	794,296	491,239
G1 Rough	KELab (KangEngine)	4096	104,009	95,889	84,954
G1 Rough	IsaacLab v2.3.2 (local)	4096	85,039	85,122	76,552

GPU-parallel simulation and training performance benchmarks.
Headless, RTX 4090 + i7-13700K; FPS counts environment control steps. Single runs with differing software stacks and simulation implementations. Benchmark details.

Features

• GPU-parallel simulation: PhysX rigid bodies and articulated robots across thousands of environments on Linux/NVIDIA.
• Robot learning: Train policies with KELab, a reinforcement-learning framework built on KangEngine with RL-Games and RSL-RL support, or use KangEngine as a simulation backend for MimicKit.
• Character motion: BVH/FBX motion visualization, SMPL/AMASS workflows, skeletal animation, and retargeting.
• Visualization: Interactive C++/OpenGL rendering for KangEngine and Newton simulation workflows.

Quick Start

Use Python 3.12 in a virtual environment:

uv venv --python python3.12
source .venv/bin/activate
uv pip install kangengine

• OS: macOS 26+ (Apple Silicon), Ubuntu 24.04 (x86-64)
• Python: 3.12
• Simulation: PhysX 5.1 CPU on macOS; PhysX 5.8 CPU/GPU on Ubuntu
• Dependencies: OpenGL 4.1+ for visualization; NVIDIA driver 580+ for CUDA 13 on Ubuntu

See Installation for details.

import kangengine as ke


class HelloApp(ke.App):
    def setup(self):
        self.materials = self.create_standard_materials()
        self.scene.add_ground("/ground", scale=10.0)
        box = self.scene.add_mesh(
            "/box",
            ke.geometry.create_cube_data(1.0),
            self.materials.pbr,
            color=ke.Vec4(0.8, 0.3, 0.02, 1.0),
        )
        box.set_local_translation(ke.Vec3(0.0, 0.5, 0.0))
        self.set_camera_view([3.0, 2.5, 4.0], [0.0, 0.5, 0.0])


app = HelloApp()
app.initialize(width=1920, height=1080, hide_ui=False, up_axis=ke.UpAxis.Y)
app.start()

See Hello App for lifecycle callbacks.

Examples

Workflow	Start here
Create a scene and manipulate objects	Hello App / First Scene
Drop a ball with physics	First Simulation
Load and simulate articulated robots	Articulation
Load and play character motion	Load Motion
Train RL policies with KELab	KELab / Benchmarks
Train RL policies with Mimickit	MimicKit
Visualize Newton simulations	Newton integration

Complete scripts are in python/examples. Clone the repository to run those files; they are separate from the installed Python package.

License

KangEngine is licensed under the MIT License. Bundled third-party components retain their own licenses.

Acknowledgments

Built with NVIDIA PhysX and OpenUSD, with inspiration from MimicKit, Isaac Lab, Newton, SAPIEN, GenoViewPython, and AI4AnimationPy.