HPP-FCL — An extension of the Flexible Collision Library

FCL was forked in 2015. Since then, a large part of the code has been rewritten or removed (for the unused and untested part). The broad phase was reintroduced by Justin Carpentier in 2022 based on the FCL version 0.7.0.

If you use HPP-FCL in your projects and research papers, we would appreciate it if you'd cite it.

New features

Compared to the original FCL library, the main new features are:

• a dedicated and efficient implementation of the GJK algorithm (we do not rely anymore on libccd)
• the support of safety margins for collision detection
• an accelerated version of collision detection à la Nesterov, which leads to increased performances (up to a factor of 2). More details are available in this paper
• the computation of a lower bound of the distance between two objects when collision checking is performed, and no collision is found
• the implementation of Python bindings for easy code prototyping
• the support of new geometries such as height fields, capsules, ellipsoids, etc.
• enhance reliability with the fix of a myriad of bugs
• efficient computation of contact points and contact patches between objects
• full support of object serialization via Boost.Serialization

This project is now used in many robotics frameworks such as Pinocchio, an open-source software that implements efficient and versatile rigid body dynamics algorithms, the Humanoid Path Planner, an open-source software for Motion and Manipulation Planning. HPP-FCL has also been recently used to develop Simple, a new (differentiable) and efficient simulator for robotics and beyond.

A high-performance library

Unlike the original FCL library, HPP-FCL implements the well-established GJK algorithm and its variants for collision detection and distance computation. These implementations lead to state-of-the-art performances, as depicted by the figures below.

On the one hand, we have benchmarked HPP-FCL against major software alternatives of the state of the art:

1. the Bullet simulator,
2. the original FCL library (used in the Drake framework),
3. the libccd library (used in MuJoCo).

The results are depicted in the following figure, which notably shows that the accelerated variants of GJK largely outperform by a large margin (from 5x up to 15x times faster). Please notice that the y-axis is in log scale.

On the other hand, why do we care about dedicated collision detection solvers like GJK for the narrow phase? Why can't we simply formulate the collision detection problem as a quadratic problem and call an off-the-shelf optimization solver like ProxQP)? Here is why.

One can observe that GJK-based approaches largely outperform solutions based on classic optimization solvers (e.g., QP solver like ProxQP), notably for large geometries composed of tens or hundreds of vertices.

Open-source projects relying on Pinocchio

• Pinocchio A fast and flexible implementation of Rigid Body Dynamics algorithms and their analytical derivatives.
• IfcOpenShell Open source IFC library and geometry engine.
• Crocoddyl A software to realize model predictive control for complex robotics platforms.
• TSID A software that implements a Task Space Inverse Dynamics QP.
• HPP A SDK that implements motion planners for humanoids and other robots.
• Jiminy A simulator based on Pinocchio.
• ocs2 A toolbox for Optimal Control for Switched Systems (OCS2)

C++ example

Both the C++ library and the python bindings can be installed as simply as conda -c conda-forge install hpp-fcl. The .so library, include files and python bindings will then be installed under $CONDA_PREFIX/lib, $CONDA_PREFIX/include and $CONDA_PREFIX/lib/python3.XX/site-packages.

Here is an example of using HPP-FCL in C++:

#include "hpp/fcl/math/transform.h"
#include "hpp/fcl/mesh_loader/loader.h"
#include "hpp/fcl/BVH/BVH_model.h"
#include "hpp/fcl/collision.h"
#include "hpp/fcl/collision_data.h"
#include <iostream>
#include <memory>

// Function to load a convex mesh from a `.obj`, `.stl` or `.dae` file.
//
// This function imports the object inside the file as a BVHModel, i.e. a point cloud
// which is hierarchically transformed into a tree of bounding volumes.
// The leaves of this tree are the individual points of the point cloud
// stored in the `.obj` file.
// This BVH can then be used for collision detection.
//
// For better computational efficiency, we sometimes prefer to work with
// the convex hull of the point cloud. This insures that the underlying object
// is convex and thus very fast collision detection algorithms such as
// GJK or EPA can be called with this object.
// Consequently, after creating the BVH structure from the point cloud, this function
// also computes its convex hull.
std::shared_ptr<hpp::fcl::ConvexBase> loadConvexMesh(const std::string& file_name) {
  hpp::fcl::NODE_TYPE bv_type = hpp::fcl::BV_AABB;
  hpp::fcl::MeshLoader loader(bv_type);
  hpp::fcl::BVHModelPtr_t bvh = loader.load(file_name);
  bvh->buildConvexHull(true, "Qt");
  return bvh->convex;
}

int main() {
  // Create the hppfcl shapes.
  // Hppfcl supports many primitive shapes: boxes, spheres, capsules, cylinders, ellipsoids, cones, planes,
  // halfspace and convex meshes (i.e. convex hulls of clouds of points).
  // It also supports BVHs (bounding volumes hierarchies), height-fields and octrees.
  std::shared_ptr<hpp::fcl::Ellipsoid> shape1 = std::make_shared<hpp::fcl::Ellipsoid>(0.7, 1.0, 0.8);
  std::shared_ptr<hpp::fcl::ConvexBase> shape2 = loadConvexMesh("../path/to/mesh/file.obj");

  // Define the shapes' placement in 3D space
  hpp::fcl::Transform3f T1;
  T1.setQuatRotation(hpp::fcl::Quaternion3f::UnitRandom());
  T1.setTranslation(hpp::fcl::Vec3f::Random());
  hpp::fcl::Transform3f T2 = hpp::fcl::Transform3f::Identity();
  T2.setQuatRotation(hpp::fcl::Quaternion3f::UnitRandom());
  T2.setTranslation(hpp::fcl::Vec3f::Random());

  // Define collision requests and results.
  //
  // The collision request allows to set parameters for the collision pair.
  // For example, we can set a positive or negative security margin.
  // If the distance between the shapes is less than the security margin, the shapes
  // will be considered in collision.
  // Setting a positive security margin can be usefull in motion planning,
  // i.e to prevent shapes from getting too close to one another.
  // In physics simulation, allowing a negative security margin may be usefull to stabilize the simulation.
  hpp::fcl::CollisionRequest col_req;
  col_req.security_margin = 1e-1;
  // A collision result stores the result of the collision test (signed distance between the shapes,
  // witness points location, normal etc.)
  hpp::fcl::CollisionResult col_res;

  // Collision call
  hpp::fcl::collide(shape1.get(), T1, shape2.get(), T2, col_req, col_res);

  // We can access the collision result once it has been populated
  std::cout << "Collision? " << col_res.isCollision() << "\n";
  if (col_res.isCollision()) {
    hpp::fcl::Contact contact = col_res.getContact(0);
    // The penetration depth does **not** take into account the security margin.
    // Consequently, the penetration depth is the true signed distance which separates the shapes.
    // To have the distance which takes into account the security margin, we can simply add the two together.
    std::cout << "Penetration depth: " << contact.penetration_depth << "\n";
    std::cout << "Distance between the shapes including the security margin: " << contact.penetration_depth + col_req.security_margin << "\n";
    std::cout << "Witness point on shape1: " << contact.nearest_points[0].transpose() << "\n";
    std::cout << "Witness point on shape2: " << contact.nearest_points[1].transpose() << "\n";
    std::cout << "Normal: " << contact.normal.transpose() << "\n";
  }

  // Before calling another collision test, it is important to clear the previous results stored in the collision result.
  col_res.clear();

  return 0;
}

Python example

Here is the C++ example from above translated in python using HPP-FCL's python bindings:

import numpy as np
import hppfcl
# Optional:
# The Pinocchio library is a rigid body algorithms library and has a handy SE3 module.
# It can be installed as simply as `conda -c conda-forge install pinocchio`.
# Installing pinocchio also installs hpp-fcl.
import pinocchio as pin

def loadConvexMesh(file_name: str):
    loader = hppfcl.MeshLoader()
    bvh: hppfcl.BVHModelBase = loader.load(file_name)
    bvh.buildConvexHull(True, "Qt")
    return bvh.convex

if __name__ == "__main__":
    # Create hppfcl shapes
    shape1 = hppfcl.Ellipsoid(0.7, 1.0, 0.8)
    shape2 = loadConvexMesh("../path/to/mesh/file.obj")

    # Define the shapes' placement in 3D space
    T1 = hppfcl.Transform3f()
    T1.setTranslation(pin.SE3.Random().translation)
    T1.setRotation(pin.SE3.Random().rotation)
    T2 = hppfcl.Transform3f();
    # Using np arrays also works
    T1.setTranslation(np.random.rand(3))
    T2.setRotation(pin.SE3.Random().rotation)

    # Define collision requests and results
    col_req = hppfcl.CollisionRequest()
    col_res = hppfcl.CollisionResult()

    # Collision call
    hppfcl.collide(shape1, T1, shape2, T2, col_req, col_res)

    # Accessing the collision result once it has been populated
    print("Is collision? ", {col_res.isCollision()})
    if col_res.isCollision():
        contact: hppfcl.Contact = col_res.getContact(0)
        print("Penetration depth: ", contact.penetration_depth)
        print("Distance between the shapes including the security margin: ", contact.penetration_depth + col_req.security_margin)
        print("Witness point shape1: ", contact.getNearestPoint1())
        print("Witness point shape2: ", contact.getNearestPoint2())
        print("Normal: ", contact.normal)

    # Before running another collision call, it is important to clear the old one
    col_res.clear()

Acknowledgments

The development of HPP-FCL is actively supported by the Gepetto team @LAAS-CNRS, the Willow team @INRIA and, to some extend, Eureka Robotics.