KangEngine

A lightweight C++/Python engine for visualizing motion, robotics assets, and PhysX-based simulation experiments.

Note: This is a personal, long-term codebase dedicated to ongoing research and self-study. It is crafted as a lifetime sandbox for exploring computer graphics (especially character animation) and robotics control.

📚 Official Documentation

Character Animation	Simulation	Rendering

KangEngine is built for quick iteration around character motion, robot assets, and simulation visualization. The C++ side owns the renderer, scene graph, asset loaders, and PhysX integration; the Python package exposes the same runtime for scripts, motion tools, control experiments, and MimicKit integration.

Requirements

• OS: macOS on Apple Silicon (Tahoe tested) or Linux (Ubuntu 24.04 tested)
• Build: C++17, CMake, vcpkg, PhysX 5.1/5.8
• Python: 3.12 for bindings and examples
• Graphics & GPU: OpenGL 4.1+ compatible GPU (NVIDIA GPU required only for PhysX GPU/CUDA workflows)

What It Does

• Multi-Format Asset Viewer: Loads FBX, BVH, MJCF, OpenUSD, OBJ, and STL assets.
• Motion Inspection: Visualizes skeletal motion, FK poses, root trajectories, contacts, and tracking targets.
• Interactive Visualization: Provides skinned character rendering, skeleton overlays, debug drawing, and scene interaction tools.
• PhysX Simulation: Runs rigid bodies, articulated robots, and motion-tracking control experiments.
• Newton Viewer: Uses KangEngine to visualize Newton simulation state.
• Python Workflows: Provides Python APIs for motion editing, IK/control experiments, simulation scripts, and MimicKit integration.

Quick Start

Build the C++ executable:

cmake --preset=vcpkg
cmake --build build/release
make run2

Build and install the Python package:

uv venv python/.venv --python 3.12
source python/.venv/bin/activate
make build_python
uv pip install -e ./python

Run a Python motion viewer:

python ./python/examples/view_bvh_character.py

Run a Python PhysX example:

python ./python/examples/sim_world_minimal.py

See Build Guide for platform setup, PhysX, USD, and Python binding details.

Feature Overview

Rendering

• OpenGL renderer with instanced mesh drawing and a lightweight graphics abstraction layer.
• SceneGraph and ExternalBuffer transform paths for authored scenes and high-throughput simulation visuals. See examples/physics/physx_h1_instancing.cpp.
• Shadow mapping, skybox rendering, gamma post-processing, ImGui tooling, and post-process selection outlines.
• GPU skinning for animated FBX characters.
• Debug rendering utilities for lines, arrows, coordinate axes, points, and camera frustums.

Asset Import

• FBX: skeletons, animation clips, static meshes, and skinned meshes.
• BVH: skeleton hierarchy, frame time, root motion, and local joint rotations.
• MJCF: articulated characters, collision geometry, joints, and inertials.
• USD: mesh traversal, material subsets, texture loading, and scene serialization(TODO). Development builds use a local OpenUSD installation; release wheels bundle the vcpkg OpenUSD runtime.
• OBJ/STL static mesh import.

Simulation & Animation

• PhysX rigid bodies and articulated robot simulation.
• Skeleton trees, sampled motion clips, FK, pose states, and skeleton visual bridges.
• Bridges for syncing physics, skeletons, and skinned characters to scene/render state.
• Experimental XPBD cloth simulation. See examples/physics/xpbd_cloth.cpp.

Advanced Runtime Paths

• ExternalBuffer visual sync: skips per-object scene graph mutation during simulation and uploads batched transforms directly to renderer-owned instance buffers. This is the preferred path for large simulation visuals such as many robot bodies or rigid objects.
• XPBD cloth: a non-PhysX cloth simulation experiment used to explore constraint-based deformable simulation.

Python

• pybind11 bindings for app, scene, animation, physics, asset, and renderer-facing APIs.
• Headless simulation and live visualization helpers.
• Motion editor modules for trajectories, contacts, targets, and tracking overlays.
• MimicKit-compatible backend adapter.

See Simulation API for the recommended KangSimWorld workflow and how it relates to lower-level PhysX wrappers.

Development Notes

This project is evolving quickly. While the main workflows are stable, some internal and high-level APIs are under active development and subject to change.

Planned work:

• WebGPU backend implementation beyond the current placeholder.

RL With MimicKit

KangEngine can be used as a backend engine of MimicKit through KangEngine's Python package. Use the backend_kangengine branch of MimicKit and keep MimicKit in a separate Python environment.

Note: GPU contact sensors provide contact count, contact state, and accumulated normal impulse and force. Tangential friction impulse and a full six-axis contact wrench are not currently exposed as sensor outputs.

engine_name: "kangengine"

control_mode: "pos"
control_freq: 30
sim_freq: 120
env_spacing: 5
enable_self_collisions: false

References

KangEngine is inspired by and built upon ideas from these excellent projects:

• MimicKit: Motion imitation and RL experiment structure.
• Isaac Lab: Robot learning workflows and simulation tooling.
• Newton: Robotics API shape and GPU simulation design.
• SAPIEN: Robotics simulation and GPU simulation design.
• GenoViewPython: Skeletal animation and motion-debug visualization.
• AI4AnimationPy: Motion modules and animation-engine structure.
• NVIDIA PhysX: Rigid body and articulation simulation.
• OpenUSD: Prim/path concepts and optional USD asset interchange.