Differentiable Collision Detection (DiffCollision)

Official implementation for Robust Differentiable Collision Detection for General Objects.

Overview

DiffCollision is a robust and efficient differentiable collision detection framework for general 3D objects, supporting both convex and concave geometries.

It builds on the high-performance Coal library for forward collision queries and integrates with PyTorch autograd for gradient computation, enabling seamless use in optimization and learning pipelines.

Key Features

• Parallel multi-mesh batching — Efficiently handles multiple mesh pairs and batches in parallel, supporting up to around 0.5 billion convex piece pairs on a single 24 GB RTX 4090 GPU.
• Unified plug-and-play API — A single PyTorch interface with multiple differentiable backends: RS1Dist(ours, recommended), RS1Dir, RS0, FD, Analytical.
• Strong performance — Achieves >90% mm-level accuracy for convex and >70% for concave objects (DexGraspNet & Objaverse) in our main optimization benchmark.

Example

import torch
from diffcollision import DiffCollision, DCMesh

# Load meshes and define mesh pairs for collision detection
meshes = [DCMesh.from_file(f"obj{i}.obj", scale=0.1) for i in range(7)]
collision_pairs = [[0, 1], [0, 3], [3, 2], [1, 6], [6, 5]]  # total 5 mesh pairs

# Initialize differentiable collision module
diffcoll = DiffCollision(meshes, collision_pairs, method="RS1Dist")

# Batched transformation matrices for each mesh, shape (batch, n_mesh, 4, 4)
transforms = torch.eye(4, requires_grad=True)[None, None].expand(13, 7, 4, 4)

# Forward & backward. wp1/wp2: witness points, shape (batch, n_pair, 3)
result = diffcoll.forward(transforms)   # including wp1/wp2, normal, sdf, etc.
((result.wp1 - result.wp2) ** 2).sum().backward()   

For more examples:

• tests/test_sphere.py — basic usage & installation verification
• tests/test_multi_mesh.py — dynamic updates of collision pairs at runtime

Installation

Create and activate a Conda environment with the coal dependency

conda create -n dcd python=3.10 coal -c conda-forge
conda activate dcd

Then install DiffCollision:

pip install -e .              # Core library only
pip install -e '.[examples]'  # Include dependencies for examples

Reproducing Paper Results

1. Prepare object assets

Download the preprocessed assets from Hugging Face:

• DGN_5k_processed.zip
• objaverse_5k_processed.zip

Unzip and organize as:

examples/assets/object
├── DGN_5k/processed_data
│   ├── core_bottle_1a7ba1f4c892e2da30711cdbdbc73924
│   └── ...
└── objaverse_5k/processed_data
    ├── 0a3d04fec6544bb697e14d0c473077ed
    └── ...

2. Run the main optimization task

CUDA_VISIBLE_DEVICES=0 bash examples/scripts/baseline/ours.sh

To debug and visualization,

python examples/main.py exp=debug n_tp=10 vis=True n_prob=1 prob_rand=None 'obj=[core_jar_5a4af043ba70f9682db9fca4b68095, sem_TissueBox_fd9c40cd2ff2aab4a843bb865a04c01a]' 'scale=[0.08238040130667001, 0.04604694495201166]' 

3. Run the grasp refinement task

python examples/grasp.py exp=grasp vis=True

License

This work is licensed under CC BY-NC 4.0.

Citation

If you find this work useful, please consider to cite:

@article{chen2025robust,
    title={Robust Differentiable Collision Detection for General Objects}, 
    author={Chen, Jiayi and Zhao, Wei and Ruan, Liangwang and Chen, Baoquan and Wang, He},
    journal={arXiv preprint arXiv:2511.06267},
    year={2025}
}