pyrosmsg

Daniel Maturana 2018, dimatura@gmail.com

What

Bidirectional pybind11 converters for a few ROS messages.

The converters will accept Python messages and convert them to C++ messages or viceversa.

This library only really makes sense for C++ code wrapped with pybind11 for use with Python.

Why

The two main languages ROS supports are C++ and Python. For any given message type, ROS will generate Python and C++ versions of that message that are interoperable when sending them over the network -- i.e., it's fine to send messages from python nodes to C++ nodes or viceversa. However, if we have a Python node that uses wrapped C++ code (something I do a lot), an issue is that (from the computer's point of view) the Python and C++ messages are actually completely unrelated objects. So e.g., to use a Python PointCloud2 message in C++ code or viceversa, conversion is needed.

One option is to use the built-in serialization capabilities of ROS - e.g., if we have a Python message, we can serialize it to a string, pass it as a string to the C++ code (trivial with pybind11), and deserialize it there. This would be the equivalent of sending the a message over the network, except we're just copying strings within the same process. This is actually implemented in this package (serialization.h). However, I found that even without the network overhead, this method is fairly slow.

Another option is to use the converters in this package. It uses pybind11 to make this conversion as transparent as possible from both the C++ and Python ends -- all supported message types are converted automatically in each direction. I should note that whenever a message crosses the Python/C++ barrier with this method, the data is copied -- there is no memory shared between the Python and C++ messages. While this will incur in some overhead, it is far less than the (de)serialization method described above.

One drawback is that the converters are (currently) written manually. It's a fairly mechanical process, and in theory it could be automated, but I have not done so yet. I would welcome converters for other message types.

How

This is a catkin ROS package, and as such should be installed in a catkin workspace like any other.

To use from the C++ side, simply #include <pyrosmsg/pyrosmsg.h>. This will register pybind11 converters that can be used to transparently accept and return the support ROS messages in your code.

To use from the Python side, simply import pyrosmsg before using wrapped functions.

So in your C++ code,

// this code is in a pybind11-wrapped module called, say, 'mycloudlib'

#include <sensor_msgs/PointCloud2.h>
#include <pyrosmsg/pyrosmsg.h>

// etc

void process_pointcloud(const sensor_msgs::PointCloud2& cloud) {
    // do fast C++ stuff with your point cloud here
}

And in your python code, presumably a ROS node,

from sensor_msgs.msg import PointCloud2
import pyrosmsg
import mycloudlib

def my_pc2_callback(pc2msg):
    # pc2msg is a *python* sensor_msgs.msg.PointCloud2
    # transparently pass the PointCloud2 to your C++ code.
    # under the hood, pyrosmsg will create a C++ copy of the message
    # and pass it to the wrapped mycloudlib function.
    mycloudlib.process_pointcloud(pc2msg)

Currently supported messages

ros::Time
std_msgs::Header
geometry_msgs::Point
geometry_msgs::Vector3
geometry_msgs::Quaternion
geometry_msgs::Transform
geometry_msgs::TransformStamped
sensor_msgs::PointField
sensor_msgs::PointCloud2
sensor_msgs::Image

Platforms

Only tested with Ubuntu 16.04, ROS Kinetic, and Python 2.7. Would probably work with other Linux and (Indigo and newer) ROS distros. Python 3, probably not without minor modifications.

TODO

It would be nice to avoid copying data, if possible. In my own experience copying image and point cloud data within a process has never been a bottleneck compared to whatever data processing I perform with the data -- after all, the copy is just a memcpy of a contiguous chunk a memory. But for massive amounts of data unnecessary copying is a drag. In theory, this should be possible -- the only difficulty is correct memory management across the Python and C++ border.

LICENSE

BSD, see LICENSE.

dimatura / pyrosmsg