🚧 This project is in active beta development. This means the API is unstable, features may be added or removed, and breaking changes are likely to occur frequently and without notice as the design is refined.

Efficient Renderer for Natural and Interplanetary Environments (Ernie)

Ernie is a physically accurate, real-time, spectral renderer for interplanetary environments.

We provide C++ and Python APIs and a Hydra-like scene API. It is a Vulkan 1.3 rasterizer that runs on Linux, and builds on Windows and macOS (experimental: not yet tested there). Ernie is designed around physical quantities rather than artistic parameters: lights, materials and cameras are described the way they are measured, so a render can be compared against a real instrument. For reinforcement learning, Ernie renders batches of cameras in one step, shares data zero copy with any DLPack tensor (PyTorch, JAX, Warp, NumPy), and has a fast mode with simpler shading.

Ernie's key features related to planetary robotics and sensing:

• Camera effects and modeling: exposure, aperture, motion blur, rolling shutter, lens distortion and vignetting, with perspective, fisheye and orthographic projections.
• Real camera specs: from quantum efficiency curves to full well, read noise, dark current, PRNU, DSNU, gain and ADC bit depth, and line-scan pushbroom cameras like LROC's, enabling users to model the noise and effects of those instruments.
• Terrain: a first-class citizen, with a built-in Terrain prim type with continuous level of detail through quadtrees, updated in place where its heights change. With a radius, a Terrain is a whole planet or moon, drawn from orbit down to the ground with cube-sphere level of detail.
• GPU interop: deforming meshes, instances and points from CUDA or Warp arrays without copies, useful for rover simulations and particle simulations.
• Spectral-first rendering: light is carried in spectral bands (16 across 380-780 nm by default) from the light sources, through the materials, to the sensor.
• Star fields: a built-in StarField prim allowing users to plug in their own star catalogs, with brightness from magnitudes and color from B-V indices, so you can see how camera exposure affects the visible stars.
• Lighting: distant lights have an angle parameter, giving smooth penumbrae based on the light's width in the sky. Lights also take illuminants (blackbody, CIE daylight) and custom spectral curves, for example to model Earthshine on the lunar surface.
• Customizable and extendable: write your own sensor filters in Slang and build up your own chains (noise models, lens flare, vignetting), designed for efficiency and throughput, and plug in custom BRDFs for the planetary reflectances you encounter. Hapke, Lommel-Seeliger, lunar-Lambert, Oren-Nayar and Lambert come built in.
• Batch rendering: many environments ("worlds"), each with its own camera, rendered in one step for RL and Monte Carlo workloads.
• Debug drawing: an easy-to-use API to add custom labels, trajectories and overlays to the scene.
• Floating origin: prevents precision issues in large-scale scenes.
• Viewer: a viewer class to easily spin up a GUI with customizable panels to integrate with your own simulations or workflows.

For a more detailed overview of the features, see our documentation.

Future work:

• A fully USD- and glTF-compliant renderer, loading OpenUSD, glTF, OBJ, STL and 3D Tiles, with the glTF material system, UsdPreviewSurface shaders and MDL.
• Atmosphere rendering, both on and above a planet, for LEO simulations or Mars rovers.
• Dust and other volumes from particle simulations through NanoVDB.
• Global illumination for reflections
• USGS CSM camera models for advanced, non-standard cameras and their noise.
• OpenXR, so users can easily connect headsets to an Ernie viewer.
• A ray-tracing backend alongside the rasterizer.
• A fast mode for RL training: a forward renderer with a smaller feature set and higher throughput.
• Windows support, and tested macOS support.
• Continued test assets and benchmark scenes.
• Guassian splat and particle fields
• More sensor support for Lidars and Radars.
• Integrations into Newton Viewer, Chrono Sensor, and IsaacLab.
• Compare benchmarks with Newton Warp Raytrace, Madrona both in speed, throughput, and image quality.

Getting Started

Python

pip install ernie-renderer

from PIL import Image

import ernie
from ernie import tokens

scene = ernie.Scene()
scene.define("/World/Ground", ernie.GroundPlane)
scene.define("/World/Rock", ernie.Sphere, translate=(0, 0, 1))  # 1 m up
scene.define("/World/Sun", ernie.DistantLight).set(
    tokens.TRANSFORM, ernie.distant_light_transform((0.4, -0.6, 0.7))
)
scene.define("/World/Camera", ernie.Camera).set(
    tokens.TRANSFORM, ernie.look_at_transform((0, -8, 3), (0, 0, 1))
)
scene.define(
    "/Render/Settings", ernie.RenderSettings, camera="/World/Camera", resolution=(640, 480)
)

renderer = ernie.Renderer()
image = renderer.render(scene)  # a (480, 640, 4) uint8 NumPy array
Image.fromarray(image).save("ernie.png")

C++

With Ernie installed (see Installation), find it with CMake:

find_package(Ernie REQUIRED)
target_link_libraries(my_app PRIVATE Ernie::ernie)

#include <ernie/scene/camera_schemas.h>
#include <ernie/scene/geometry_schemas.h>
#include <ernie/scene/light_schemas.h>
#include <ernie/scene/render_schemas.h>
#include <ernie/scene/schema.h>
#include <ernie/scene/tokens.h>
#include <ernie/scene/xform.h>
#include <ernie/session/renderer.h>

int main() {
    using namespace ernie;
    Scene scene;
    GroundPlane::Define(scene, "/World/Ground");
    Sphere::Define(scene, "/World/Rock");
    SetXform(scene, "/World/Rock", {.translate = glm::dvec3(0.0, 0.0, 1.0)});  // 1 m up
    DistantLight::Define(scene, "/World/Sun");
    scene.Set("/World/Sun", tokens::kTransform, DistantLightTransform({0.4, -0.6, 0.7}));
    Camera::Define(scene, "/World/Camera");
    scene.Set("/World/Camera", tokens::kTransform,
              LookAtTransform({0.0, -8.0, 3.0}, {0.0, 0.0, 1.0}, {0.0, 0.0, 1.0}));
    RenderSettings::Define(scene, "/Render/Settings")
        .SetCamera("/World/Camera")
        .SetResolution({640, 480});

    Renderer renderer;
    Tensor image = renderer.Render(scene);  // 480 x 640 x 4 bytes of LdrColor on the CPU
}

For the prim types and their attributes, transforms, relationships, arrays, validation, loading scenes from JSON and declaring your own types, see Describing scenes.

Examples

Planned future examples.

Cameras

• LRO NAC and WAC
• Shadowcam
• Mars2020 Navcam, Mastcam
• LWIR, NIR, UV
• Pinhole, fisheye, orthographic
• Narrowband camera
• Distortion

Camera Filters and Noise

• Vignetting
• Lens flare
• Bloom
• Narrowband filter viewing laser lines

Terrain

• SLDEM global and multiple tiles
• Mars terrains
• Asteroid shape files
• Saturn and its rings
• Pluto

Lighting

• Distant light
• Custom radiance curves to model Earthshine
• Dome light shadows
• Illuminants across the same scene
• Laser lines competing with sunlight

Star Field

• Hipparcos star catalog
• Stars viewed from another galaxy

Atmosphere

• Earth atmosphere (from orbit and ground)
• Mars atmosphere

Batched and RL

• Batched camera rendering and setting poses

Viewer

• Adding a custom panel
• Hiding default panels and menus
• Extending HUD options
• Physics debug draw overlays

Installation

Precompiled binaries

The Python wheel is on PyPI as ernie-renderer (pip install ernie-renderer, then import ernie; Python 3.12+). It bundles everything, the viewer included, but a Vulkan driver.

Each GitHub release also has an ernie package for Ubuntu 22.04+, Debian 12+ and RHEL 9+ (and a tarball for other distributions), with the library, the viewer, and the headers and CMake package for building against it.

sudo apt install ./ernie_*.deb    # Debian, Ubuntu
sudo dnf install ./ernie-*.rpm    # RHEL, Rocky, Fedora

Both need a Vulkan 1.3 driver: your GPU vendor's, or Mesa's (mesa-vulkan-drivers, which includes the lavapipe software driver for machines without a GPU).

Build from source

Ernie needs CMake 3.24+, a C++20 compiler and a Vulkan 1.3 driver; it fetches its other dependencies (see third_party/README.md).

cmake -S . -B build -DERNIE_BUILD_TESTS=ON -DERNIE_INSTALL=ON -DERNIE_BUILD_PYTHON=ON -DERNIE_BUILD_APP=ON
cmake --build build
ctest --test-dir build -j4              # C++, GPU and Python tests
build/bin/ernie                         # the viewer
cmake --install build --prefix ~/.local # libernie, headers and the CMake package

A Python wheel bundles the module with the Slang runtime it loads:

uv build --wheel --out-dir dist .
uv pip install dist/ernie_renderer-*.whl

Packages are built with cmake --preset package-deb (or package-rpm), cmake --build --preset package-deb and cpack --preset package-deb.

See CONTRIBUTING.md for the build options, testing without a GPU on lavapipe, and benchmarks.

License

Ernie is licensed under the BSD 3-Clause License. The third-party libraries it bundles keep their own licenses, listed in third_party/README.md.