[ros-users] The future of ROS 2.0 protocol changes

Aaron Sims aaron at happyartist.net
Sun Sep 14 20:40:17 UTC 2014

Dear ROS Teams, Thanks for an informative time at ROSCon 2014. I had to leave about an hour and a half early to catch a flight, and I was left with my head spinning about the the decision to change protocols in ROS 2.0. This is an important decision, and I will share my concerns, as well as my thoughts on a direction I personally would of taken with ROS 2.0, and the protocol directions. This topic seems better fit for a blog, however, an open discussion is important.  First let me say the direction of ROS 2.0 makes sense, the complete shift to ROS DDS protocol does not seem wise. There are many good reasons to implement DDS in ROS, however, the way it is implemented can make or break ROS in the future.  Many well intended protocol implementations end up at dead ends. Two examples are:  HTTP NG (Next Generation)  http://www.w3.org/Protocols/HTTP-NG/Activity.html Gnutella 2 http://en.wikipedia.org/wiki/Gnutella2 I'd like to share the approach of the Happy
 Artist RMDMIARCSM (Robot Control System Messenger) API. The RCSM implements a plugin architecture that allows multiple robot operating system clients to be registered/accessed, run simultaneously, and inter operate together, or simply to plugin a single client such as the Happy Artist ROS Client for TCPROS/UDPROS.  As a ROS user this allows any protocol to plugin, and work from any 3rd party vendor DDS implementation, TCPROS/UDPROS, or any other protocol. In the Happy Artist RMDMIA implementation the RCSM component serves as an interface to the rest of the system so that protocols can change, while code written for autonomous/user control follows the same pattern, allowing interchangeability, and forward/backward compatibility between other aspects of the system. If ROS were to approach 2.0 like this, libraries like MoveIt (an example of a library name (no idea if it is tied to TCPROS/UDPROS protocols in ROS 1.0 implementation), could run on any
 communication protocol without ever needing to know anything about the protocols it was operating on. This approach needs to be applied to all areas of the system so that a user of any component can interchange their libraries to ROS 2.0 for 100% forward, and backward compatibility. I am personally concerned that researchers may find themselves using ROS alternatives due to the time/money that will be lost while they wait for a viable ROS 2.0 solution. Why would a researcher/developer invest resources into a platform that was going to completely change nullifying their personal investment in that platform? It is evident ROS 2.0 needs support similar to ACID transactions for autonomous robotics that the DDS implementation will support (primarily for people safety, and system stability of commercialized autonomous systems).  It is a bad, bad idea to throw the current ROS users under the bus in the interim. The implementation architecture direction for ROS
 2.0 I am suggesting alleviates this scenario, and allows ROS to backtrack a little with its users to continue improving ROS 1.0 functionality in TCPROS, and UDPROS.
In the interim here are some ROS 1.0 improvements that could greatly improve ROS performance with substantial latency decrease, while increasing data bandwidth:
Add support for UDPROS on every ROS supported client, and server with a few minor enhancements. (Some of these suggestions came out of the BOF at ROSCon discussion on latency and performance, and some I have been attempting to subtly hint at with multiple questions on ROS Answers over the last year). UDP Jumbogram support:This is a very simple enhancement. Allow unbounded UDP Datagram sizes, and during protocol handshake (XMLRPC negotiation in UDPROS), and add a flag to the publisher configuration that says allow UDP jumbogram support, or specify the maximum datagram size. ROS currently puts their datagram size limit around 1500 bytes. If the system hardware supports jumbograms, the only change is supporting a configurable maximum datagram size.  UDP Multicast support:Adding a header attribute called multicast could allow a topic or service (not aware service is currently supported for UDPROS, but it should be) to configure the associated Publisher to
 perform a broadcast (A concern of multiple subscribers to a topic causing a major problem with system performance can be alleviated with a feature request by a person in the BOF get together for specifying the numeric limit to subscribers of a particular topic).  Note: (Based on my personal UDPROS implementation experience, adding Multicast support would be fairly simple to add into the Happy Artist Java client, so I assume the same might apply on other platforms that have UDPROS support).  Maximum Topic Subscribers support:Allow the the system designer to specify the maximum number of subscribers on any given topic. The use case given was a high bandwidth topic that multiple users subscribe to and cause the system to hang. The user wanted to implement a UDPROS topic with Multicast support for a single subscriber. Note: Based on my personal knowledge of the CPP client (somewhat limited), the client uses low level arrays rather than Data Collection
 Objects, therefore the incremental nature of arrays in CPP code should make this task fairly straight forward to implement. If you have any questions or would like to have further discussion, I look forward to discussion.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ros.org/pipermail/ros-users/attachments/20140914/0f779b6a/attachment-0001.html>

More information about the ros-users mailing list