You are right that it's usually more ideal to have just one web socket connection rather than a connection for each topic. I was accepting it as an unfortunate downside of talking to ROS nodes, but now that you've raised the point, it seems to be more serious of a downside than I thought. Specifically, it'd be difficult to do things like implement access control for each topic, which is particularly vital for being on the web.

After looking into this a little more, however, it looks like whichever DDS system is used, if there is support for some sort of shared memory communication then having a bridge shouldn't be a big performance hit anyway, especially if the marshaling format ends up being JSON.

Since it's all contingent on shared memory access, is OSRF leaning towards a particular implementation of DDS that has this feature?

I do agree with Chad though. It's rather bold to assume that most of the best feedback for ROS 2.0 will be from people that are interested enough to join yet another mailing list, especially since this one is so relevant. From the earliest days, ROS's design has been informed and improved by constant interactions with roboticists. For the benefit of ROS 2.0, I think this process should continue since there is a very real possibility that bad design decisions will result in developers just continuing to use ROS 1.x or even worse, not updating their libraries to support it.

On Tue, Sep 16, 2014 at 5:13 PM, William Woodall <william@osrfoundation.org> wrote:
On Tue, Sep 16, 2014 at 3:17 PM, Hai Nguyen <haidai@gmail.com> wrote:
I am not a WebSocket expert but from my understanding, HTTP is used just for the initial handshake so high frequency tf messages should work.

You're right, I was thinking the actual transport was over HTTP, but it is only for the handshaking, I suppose I confused this detail with how REST works.
 
And although I don't have measurements either, the current process where we encode to ROS, send, decode from ROS, encode to JSON, send, and finally decode from JSON is certainly much slower (high latency) and processor intensive than if messages started out as JSON. You also don't strictly need to send JSON at the server end—not that it's that much different from ROS messages—as WebSockets doesn't define a serialization format. For example, you can always send just binary blobs.

As Bob Dean suggested we could make it so that the message is deserialized directly into JSON (we want to support other in memory formats like OpenCV and PCL types as well). This would make the bridge more efficient. But I think the websockets need to still all go over 80 correct? meaning that you need a single point of entry to a remote machine. It seems like a more natural architecture to have one process serve as the gateway to a ROS system for the browser.
 

I agree that having WebSockets in each node, in addition to existing protocols, would be nice but perhaps complex. From what I can see with that linked DDS document, they're just talking about another bridge solution and that is the solution that we have now, which is kind of terrible. Although I didn't expect much since OMG, as you put it, is "slow to adapt to changes and therefore arguably doesn’t always keep up with the latest trends in software engineering." It looks like, for OMG, the web falls into this category of "latest trend in software engineering" that they don't "always keep up with." 

Well the fact that they have been discussing it for over two years even though websockets are only about three years old, and that they have a beta specification out is pretty good for a standards body, IMO. And I have yet to see a technical alternative to the bridge design for communicating with a middleware from a browser.
 

Personally, I feel like maybe having WebSockets be specifiable as a transport, instead of TCP or more exotic protocols, would seem like a good compromise.

Websockets are over TCP, I'm not sure what you mean here. I don't think it is a good idea to have multiple wire protocols for ROS officially. I think in this case having a bridge into another system like websockets is the right thing to do, but I might be convinced otherwise if we come up with a compelling technical solution which involves this.
 

Totally not related, but I've just (re)noticed that rosmaster and the parameter server runs on XMLRPC, which is a whole lot simpler to use and more sophisticated than ROS services (supporting things like exceptions), so why does ROS services need to be there at all? I feel like the master, in ROS 2.0, should hand out plain xmlrpc connections to services when requested by name.

I've seen that XMLRPC is not suitable for high frequency service requests and because it relies on XML it is difficult if not impossible to implement on embedded systems (see rosc's difficulties with XML parsing). Also ROS Services report errors too. In ROS 2.0 we will likely use DDS's RPC system once it becomes available, allowing us to keep one dependency for both pub/sub and services. Since ROS 2.0 will be using DDS, there will be no ROS master, DDS has a distributed discovery system which uses UDP multicast. This is another reason a bridge is technically necessary to communicate with a browser, since the browser cannot discover it's peers in the ROS network without a proxy in the ROS network (the bridge).
 

On Tue, Sep 16, 2014 at 2:38 PM, William Woodall <william@osrfoundation.org> wrote:
I think the limiting factor here is that browsers only support websockets as a means to communicate with other processes (without browser specific extensions). The overhead of doing socket communication over HTTP leads to the difficulty with high frequency messages like tf. I'll try to address some other things inline.

On Tue, Sep 16, 2014 at 1:01 PM, Hai Nguyen <haidai@gmail.com> wrote:
I'm not particularly familiar with DDS, but does it work nicely with browsers? either through some sort of translation or better yet, natively?

People have been using DDS with websockets:


There is a DDS-web specification "in progress", the beta of the spec was released in November 2013:


Briefly looking at it, they plan to do what we do, which is to run a native DDS program and have access to data from the browser go through that program to get data using websockets.
 
I'm interested as the current ROS to WebSockets bridge is particularly ugly: the bridge has to subscribe to all the messages that any web client would need to listen to and then rebroadcast them, which introduces additional delays making it horribly painful to use for things like teleop with large messages like images or point clouds.

I'm not convinced the rebroadcast is the core performance issue here, but you could imagine a system where every node has a websocket server, and then the rosbridge node acted as a lookup service for the web browser, allowing it to connect directly to any node. However, I'm not convinced this is a better solution because it adds a lot of complexity. I really think that the over head of converting to JSON and then sending data over HTTP using websockets is the main performance problem, but I could be wrong about that since I don't have any anecdotal evidence to support it. If that is the case though, there is nothing to be done about it, since the only portable way that I know of to communicate with web browsers is through websockets using JSON.
 

From my perspective, interest, especially commercially, in getting ROS to work with the web have only grown over time. There has been push by Bosch and Brown for a while, and then Willow joined in, right before it closed shop, to build a complementary web toolkit [1] for ROS. Savioke, from their job postings, seem to be doing something webby behind the scenes too. And even for hobby/research projects, it's just so much easier to access robots over a browser compared to with Ubuntu/RViz. The fact that you get iOS and Android support, no ROS java [2] needed, almost for free through their browsers is just fantastic (and I suggest trying this if you haven't already).

All of that is true, I don't think there is anything better DDS or ROS could do from a technical perspective though. I'm open to ideas on that front, it might be good to bring up with other robot web tools people.
 

I might have missed it in my superficial lurking, but I haven't seen this issue of communicating with web clients raised with any seriousness yet. It would be a big missed opportunity if ROS 2.0 only supports talking to browsers at the level that it does now. Most projects are moving to the web. These days even my humble ipython runs a full-fledged web server in its default installation. Certainly, in the future, I feel like this will become a much more pressing issue.

This is certainly true, but in my opinion the best way to support this is to have things like a native node.js client for ROS, which should be easy since it is extended using C++. Which gives you access to all the sophisticated web service development tools with a native connection to the ROS nodes.

ipython also using websockets just like rosbridge, there is nothing that I can see that they do fundamentally different from how rosbridge does things.
 

--
You received this message because you are subscribed to the Google Groups "ROS SIG NG ROS" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-ng-ros+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.



--
William Woodall
ROS Development Team

--
You received this message because you are subscribed to a topic in the Google Groups "ROS SIG NG ROS" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ros-sig-ng-ros/BfvcjD6Rsg0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ros-sig-ng-ros+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.



--

--
You received this message because you are subscribed to the Google Groups "ROS SIG NG ROS" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-ng-ros+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.



--
William Woodall
ROS Development Team

--
You received this message because you are subscribed to a topic in the Google Groups "ROS SIG NG ROS" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ros-sig-ng-ros/BfvcjD6Rsg0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ros-sig-ng-ros+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.



--