At Shadow we had 2 different approaches: we built 2 different ROS interfaces, one for the ethercat hand we plugged to the pr2 and another one for our existing CAN hand.
I'm more familiar with the CAN interface. In this interface, we built a ROS interface on top of our existing driver. There's one publisher and one subscriber for the hand. The ROS interface reads data and publishes them periodically to the joint_states topic. It also publishes hardware information on the diagnostics topic. I also defined a couple of specific messages because it was easier in certain cases than to use the existing ROS messages. The subscriber is used to send command to the hand.
The ethercat interface is really different. It was build directly for ROS, using the pr2_controllers, etc... There's one publisher and one subscriber per joint. If I'm not mistaking, it only uses standard ROS/pr2 messages.
In both cases, one of the important thing was to build the urdf model of the robot. The model is used throughout the code and also for visualization / simulation purposes.
Hope this help, otherwise don't hesitate to recontact me if you wanted more info.
Cheers,
Ugo
On 02/08/10 07:29, Adrian wrote:
Morgan,
Thanks, that is quite helpful. The point of my question was that I want to do this in a way which would make the arm as easy as possible for existing ROS users like, for example, those familiar with the PR-2 software, to just "drop in" my hand/arm. So I would not want to go create a buch of new messages and services and such if I could help it, I would like to figure out what level of abstraction is the correct place for my code to meet up with existing code. Is the "joint space controler" the place to do this?
If so, what exactly is the "joint space controler"? From what I read in the wiki, it looks like controlers are not exactly nodes, but some kind of module which is loaded into some general real-time layer which handles all controlers on the robot. Is that correct? (I am learning about ROS as I start this, so I appreciate your patience with what probably look like dumb questions.)
Really, what I am getting at is, if someone walked up to someone and said "hi, I made a cool hand, want to use it in your research?", what would be the best case for them, as ROS users, and possibly PR-2 owners, in terms of code I could have implemented to make the hand easy to use?
I gather that the Shadow Hand folks had this exact situation, what did they do?
I'm sure others will chime in here, but IMHO the greatest thing about
ROS is that you can do whatever you want :) it's designed from the ground up to be flexible and extensible, and to not favor one approach over another equally-valid approach.
That being said, when I've built my own sensors and arms, if the
bandwidth isn't prohibitive, I first wrap up the hardware details in a ROS node that performs whatever bit-mashing is necessary to publish and subscribe to abstracted versions of what the hardware deals with,
e.g., translating to/from some packet format that contains sensor readings and actuator commands that goes over usb/ethernet/etc. to an embedded microcontroller. This node doesn't do anything beyond munging the embedded protocol into ROS messages, and maybe do graceful
startup/shutdown of the hardware. When using the real hardware, this program is started once during a session and left up for a few hours while hacking higher layers.
Once the driver wrapper is done, you can write higher layers that
either implement your own stuff or interface to other ROS software (e.g., rviz, robot state publisher, nav stack, etc. etc.). Since those higher layers aren't tied to hardware, you can use rosbag to record/play real data to the rest of the stack, and/or write a
simulation layer to swap with the "real" driver wrapper.
we just did a master-slave "puppeteer" demo of our arm that had a straightforward stack:
* human motion capture * inverse kinematics via OROCOS/KDL
* robot state publisher (forward kinematics) * joint-space controller * state estimation from encoders, accelerometers, etc. * device driver wrappers
each layer is a separate node. All of them talk to rviz in one way or
another to ease debugging. The robot state publisher (from the WG robot_model stack) converts joint angles and the robot's kinematic description (URDF) into tf frames. When doing any of this stuff, rviz is extremely helpful when converting a new robot's cad models into
URDF and in general for making sense of all the coordinate systems.
HTH, -mq
On Sun, Aug 1, 2010 at 8:34 PM, Adrian <adrian.kaehler@gmail.com> wrote: > > I am interested to know how one goes about creating an interface for
> a device, such as an arm or hand. In particular, the question is what is > the lowest level component in the ROS architecture which one would implement > first? Presumably, once that low level element was in place, one could then
> go on to create services or actions that make use of that component. > Could someone point me to (or write?) a summary of what this stack looks > like for something like the PR2 or Shadow arms? >
>