Xamla Egomo Sensorhead for UR5 + Robotiq FT-Sensor & Gripper

xamla_andreas

Hi fellow roboticists,

My name is Andreas and I am co-founder of the adaptive robotics start-up Xamla from Münster, Germany. We develop hard- and software for vision and force-torque controlled applications. For the combination UR5 + Robotiq force-torque sensor and gripper we are currently developing an open-source wireless (and low-cost) sensor head named "egomo".

The main goal of egomo is to give researches and developers a quick-start into the field of adaptive robotics. We try to lower many of the hurdles normally faced today, like high hardware costs, long time for integration, training, adapting, etc.

Here is a visualization of the sensor head mounted to a UR5:






Our first iteration of egomo is designed to work on a UR5 in combination with a Robotiq 2-finger-85 gripper, and optionally the Robotiq FT-300 force-torque sensor. Both UR and Robotiq have made affordability, flexibility and ease of integration part of their key features, so it was a perfect platform for us to start developing for!

With its Logitech C920 HD webcam, the Structure IO Depth Sensor, and an IR line laser diode, "egomo" can create 2D & 3D images of the surroundings. In addition to the Robotiq FT-Sensor we have a IMU to capture acceleration data on our custom IO board which also talks directly via MOD-Bus to the Robotiq components. A Raspberry PI3 with Ubuntu and ROS installed on it is used as the wireless relay station between sensors and an external controller PC (with GPU). All software parts, e.g. the SD-card image and ROS-nodes will be made available open-source in the time frame of Automatica 2016.

You can find a full list of egomo's technical specs on this website, and if you have any questions about specific parts, techniques used, etc., please feel free to ask!

If you want to see egomo in action, we posted this simple Pick & Place demo (finding, gripping, and stacking duplo bricks) a couple of days ago. The pose of the Duplo bricks is first coarsely measured using a point cloud obtained from a Structure IO depth sensor and subsequently refined using high-resolution stereo triangulation (the demo is running through the steps slowly for presentation purposes, we'll turn up the speed a bit for future videos).

By designing the sensor head to communicate wirelessly, we retain most of the movement potential of the robot arm – something that can quickly become a problem with standard cables. From earlier feedback we know that wireless connections are not acceptable in (some) industrial settings - for such cases it is possible to connect the Egomo sesorhead with a single high-quality Ethernet-cable (which is still better than having 4+ cables).

We are working tirelessly towards getting ready for the initial release, in the meantime we would very much welcome any constructive feedback, questions, ideas, etc. Feel free to post here, or send me an email ([email protected]). We are also looking for Beta Testers (who have some experience with ROS) – If you would like to apply, please get in touch with me via email.

Thank you very much, and looking forward to hearing your feedback!

Andreas

Samuel_Bouchard

@Vincent @Nicolas_Lauzier @Vincent_Paquin @JeanPhilippe_Mercier : have a look at this!

Alexandre_Pare

@xamla_andreas : this is a really nice product. I like the design. Can't wait to see other videos

Nicolas_Lauzier

This is amazing! Keep on the good work Andreas. I'm looking forward to see everything that people will do with such a nice integration!

xamla_andreas

@Alexandre_Pare here is another nice "unordered item removal demo": https://www.youtube.com/watch?v=YQuX608miFU (not listed on youtube, internal team video).

xamla_andreas

Here is another video that show Duplo stacking on a freely positioned target:
https://www.youtube.com/watch?v=9rsXQrF348I&feature=youtu.be

dhenderson

Andreas,

Great demo.  Question for you: It appears in your first video that you are doing some manual separation prior to running.  Do the parts have to be singulated like this or have you also been able to pick them when they aren't separated?

xamla_andreas

@dhenderson The item-removal demo uses a very simple euclidean clustering of the point-cloud. No advanced scene understanding and object classification / model-registration is done. It is too early to disclose the details of our ROSVita programming system - all I can say for now is that machine learning will be at its heart and from an AI standpoint the technique will be more sophisticated/intelligent.
Nevertheless even the most basic euclidean clustering approach has cleaned the table in our lab countless times with even very obscure items on the table: As long as a gripping-failure slightly modifies the scene sooner or later all objects are removed.

See this video (also unlisted, internal) which shows some situations where the clusters are first wrong because objects are too close to each other: https://www.youtube.com/watch?v=YQuX608miFU

Catherine_Bernier

@xamla_andreas Thanks a lot for your great demos.

I just noticed that when the gripper do not grab an object, the robot still proceed as if. It would be interesting to use the built-in feature of the gripper for object detection.

Here is a support video we made this week: Click Here!

Let me know what you think.

xamla_andreas

@Catherine_Bernier You are right, that demo did not incorporate any gripper feedback. Actually it is 'my fault' since I had no time yet to complete the ROS actionlib integration which deals with long-running asynchronous operations (as a gripper command). From a technical perspective all data is already available on the RPi... Last week we were working on the adaptive path generation which was one of our biggest issues with the ROS path-planning package MoveIt! - e.g. you see many unnecessary reconfigurations of the UR5 arm. Fortunately with some modifications from us MoveIt! now finds the shortest path with a high probability ... we will spend some time on the velocity and acceleration profile and we are quite optimistic that in simple settings as the table scene with lots of free joint-space all generated path (that take collision objects like the walls and the table into account) will be as optimal as hand-crafted solutions...

xamla_andreas

We released the xamla_egomo ros package on github. This repo includes all 3D models to print the parts of an Egomo sensor-head and the software to operate RGB + depth camera nodes + Robotiq 2-Finger-85 Gripper + Robotiq FT-300 sensor wirelessly. It also comes with lua client tools and the source code of the table clearing and duplo stacking demos. A high-speed booting RPi 3 linux image with ros will be released during the next days (also very handy for all kinds of ROS projects with the RPi that need a read-only file system). If you are interested in getting a pre-assembled Egomo device or the Xamla IO-Board for your adaptive robotics experiments please contact me. :-)

matthewd92

@xamla_andreas This is very interesting and really great work.  I sent you a PM as well regarding this.