PyBatchRender

NOTE: The demo scripts are currently buggy. They’ll be updated and uploaded to Colab in the coming months, along with new examples.

Utilities to build fast-rendered 3D environments with Panda3D and TorchRL. Includes a lightweight renderer, camera/light helpers, and example environments/demos.

Features

• Fast multi-scene rendering with a tiled offscreen buffer
• Simple helpers to add nodes, cameras, and lights
• Torch/TorchRL-friendly image tensors (uint8, BCHW)
• Optional image saving via Pillow/ImageIO/Matplotlib
• Shaders shipped with the package

Requirements

• Python >= 3.10
• Dependencies (installed automatically): panda3d, torch, tensordict, torchrl
• Optional for saving images: Pillow, imageio, matplotlib

Installation

• From PyPI (once published):

pip install -U pip
pip install pybatchrender

• From source (this repo):

pip install -U pip
pip install .
# or editable
pip install -e .

• With optional image-saving extras:

pip install "pybatchrender[images]"

• With optional CUDA interop extras (Linux/Windows with NVIDIA):

# CUDA 12.x systems (default)
pip install "pybatchrender[cuda]"

# If you are on CUDA 11.x, install the matching CuPy wheel first, e.g.:
pip install cupy-cuda11x
pip install "pybatchrender[cuda]"

Notes:

• CUDA extras are skipped on macOS automatically.
• GPU grabbing requires an NVIDIA GPU and driver with CUDA, plus OpenGL.

CUDA prerequisites (Ubuntu)

Detect your NVIDIA driver version, then install matching GL libs:

sudo apt-get update
DRV_MAJOR=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader | head -n1 | cut -d. -f1)
sudo apt-get install -y --no-install-recommends "libnvidia-gl-$DRV_MAJOR"

If needed, specify explicitly or fallback to basic EGL libs:

sudo apt-get install -y --no-install-recommends libnvidia-gl-{DRIVER_VERSION}
sudo apt-get install -y --no-install-recommends libegl1

Notes: Windows is not tested yet. Tested on macOS (Metal; no CUDA interop) and Ubuntu with NVIDIA.

Quickstart

Render a small tiled batch offscreen and get a Torch tensor back as [B, C, H, W].

from pybatchrender import PBRConfig
from pybatchrender import PBRRenderer

cfg = PBRConfig(num_scenes=4, tile_resolution=(64, 64), offscreen=True, report_fps=False)
renderer = PBRRenderer(cfg)

# Minimal setup
renderer.add_camera()
renderer.add_light()
renderer.add_node('models/box', instances_per_scene=1, shared_across_scenes=True)
renderer.setup_environment()

# Render one frame and get a BCHW tensor (uint8)
img_batch = renderer.step(return_pixels=True)
print(img_batch.shape)  # torch.Size([4, 3, 64, 64])

Demos

The package ships a few demos you can run directly:

python -m pybatchrender.demo_cartpole
python -m pybatchrender.demo_cartpole_with_logic
python -m pybatchrender.demo_many_cubes

Notes:

• Demos use offscreen mode by default; they print FPS and can optionally save images.
• Demo models like models/box and models/smiley are standard Panda3D sample models.

Using with TorchRL

pybatchrender.env.PBREnv provides a base class for TorchRL environments and utilities to define specs and render batches. See pybatchrender/demo_cartpole_with_logic.py for a full example environment with rendering.

Package data

Shader sources under pybatchrender/shaders/ are included automatically; no path configuration is needed.

Development

• Build locally:

python -m build

• Run tests/demos from the repo root to validate rendering:

python -m pybatchrender.demo_cartpole

License

MIT. See LICENSE.