RoboCup@Work League: Interview with Christoph Steup

RoboCup@Work League groups on the occasion in Brazil.

RoboCup is a world scientific initiative with the purpose of advancing the state-of-the-art of clever robots, AI and automation. The annual RoboCup occasion, the place groups collect from throughout the globe to participate in competitions throughout quite a lot of leagues, this yr occurred in Salvador, Brazil from 15-21 July. In a sequence of interviews, we’ve been assembly among the RoboCup trustees, committee members, and members, to seek out out extra about their respective leagues. Christoph Steup is an Government Committee member and oversees the @Work League. Forward of the occasion in Brazil, we spoke to Christoph to seek out out extra in regards to the @Work League, the duties that groups want to finish, and future plans for the League.

May you begin by giving us an introduction to the @Work league?

The @Work League, together with the Logistics League, kinds the Industrial League. Our purpose is to imitate among the elements of commercial manufacturing methods. An vital side of that is manufacturing facility automization and attempting to imitate the manufacturing facility of the long run, the place you have got autonomous robots constructing merchandise in accordance with buyer design. In these factories of the long run, a single piece can be produced individually for every buyer. Factories these days have huge conveyor belts and quite a lot of automization, with the duties principally achieved in the identical means, and you’ll solely construct stuff effectively for those who construct thousands and thousands of things. We’re engaged on constructing particular person items, the place automization remains to be attainable, and even a single piece will be constructed successfully. However clearly, in our RoboCup competitions, we’re not concerned about constructing on a manufacturing facility scale – we’re doing it on a really small scale. Which means our robots are sometimes 80 centimeters lengthy, the biggest are round 70 centimeters broad, and a few of them are additionally 80 centimeters excessive. So let’s say they slot in a one metre cubed field. Additionally, all our operations are achieved on the bottom. That is only for simplification as a result of constructing huge tables to make it extra life like would additionally enhance the price for RoboCup and wouldn’t give a lot extra worth.

What our robots must do is transport objects from totally different workstations. So we’ve a default configuration the place the world begins, and there are workstations with objects mendacity on them, and a few of these objects have to be transported to different workstations. The robotic wants to try this utterly autonomously. So this is among the particular issues in regards to the @Work League, that it’s utterly autonomous and there’s solely a single restart allowed per staff. So meaning the robotic actually must be dependable. One of many huge variations between medium groups and excellent groups is that the excellent groups carry out effectively on a regular basis whereas the medium groups have some good runs and a few dangerous runs.

In addition to the article transportation that you simply talked about, and there different duties that the groups want to hold out?

There are some particular duties in our league, just like the precision placement job the place the robotic wants to suit an object right into a cavity that’s primarily the identical form and measurement as the article. It’s a little bit bit like the sport that infants do to coach their dexterity.

We even have a job that’s impressed by a conveyor belt, however we’re utilizing a desk that’s continually turning. The robots want to understand stuff whereas the desk is popping. This seems a little bit bit foolish as a result of nobody would really put a rotating desk in a manufacturing facility, nonetheless that is our means of truly mimicking a conveyor belt. The conveyor belt itself can be actually, actually troublesome to combine into the competitors, so we simply abstracted that and use this rotating desk to truly have the identical problem however in a extra manageable means. And it’s nonetheless a really, very laborious problem.

Then there are some particular challenges that we combine. For instance, that robots must report their state again in order that we are able to observe what the robotic is doing. We even have a problem the place people are within the loop. For instance, the robotic brings items to a sure workstation the place a human is current, the human assembles the items, after which the human wants to offer an indication to the robotic, after which the robotic will take the piece away and put it elsewhere. That is designed to actually mimic the automated manufacturing facility circulate.

Prior to now we additionally had a problem the place the robots wanted to open a drawer, take one thing out after which shut the drawer once more. We’ve additionally had duties the place the robotic has to deal with fragile objects, like sweets, the place the robotic actually wanted to watch out in manipulating them. So generally, what differentiates us most from the Logistics League is that we’re focusing quite a bit on manipulation and all of the difficulties that include manipulation and unknown objects, whereas Logistics is extra tailor-made in direction of large-scale logistics processes with all their optimization and planning.

I used to be fortunate sufficient to attend RoboCup final yr in Eindhoven and what the groups have been doing was actually spectacular. It was additionally attention-grabbing to see how various the robots have been, and the way groups have been approaching the duties in distinct methods, with totally different grabbers and so forth.

Sure, this distinction in approaches is said to the historical past of our League, which is a little bit bit just like the Logistics League. The Logistics League was initially a sponsored demonstration by Festo, which is a big firm from Germany that creates instruments, however in addition they have a didactics space the place they supply instruments to assist folks perceive manufacturing facility optimization. The @Work League was sponsored by Kuka, the robotics firm, and, at first, they required each staff to compete with the Kuka youBot. So this was just about the default platform for our league, however in some unspecified time in the future Kuka dropped from a sponsor to only an advisor to the league, and these days they aren’t a part of the league in any respect. So when the Kuka youBot was going out of fee, the groups looked for alternate options and now we’re introduced with all kinds of robots which might be competing within the league, which I personally discover actually cool. Now we’ve all these totally different robots, all these totally different approaches, and a few work higher in some eventualities and worse in others. So we actually have a scientific method to the issue and we’re actually getting some insights into how one can sort out this drawback on a number of ranges.

Have you ever observed that among the challenges particularly are harder generally for all of the groups?

In 2018 we launched a problem of so-called arbitrary surfaces that are surfaces unknown to the groups which might be placed on high of the workstations. The groups want to have the ability to take care of these surfaces. There are two surfaces which might be actually, actually terrible for the groups – one is grass that we just about stole from the soccer competitions! We simply thought it might be humorous to strive it, and it was a very attention-grabbing drawback, particularly for among the grippers of the totally different groups. For instance, the present world champion, they’ve a inflexible gripper to allow them to have drive suggestions after they grasp. Nevertheless, the grass is basically troublesome for them – due to their rigidity, they all the time grasp the grass itself after which they pull up the grass with the article. And this led to some attention-grabbing issues, like they’re transporting the precise floor round and never solely the article. This isn’t an issue for groups utilizing a versatile gripper. Nevertheless, alternatively, the versatile gripper makes it actually troublesome to evaluate you probably have grasped the article as a result of you have got very dangerous drive suggestions. So there are two totally different approaches which have their professionals and cons in numerous eventualities.

Three robots from the @Work competitors in Brazil.

Are you introducing any new duties for this yr?

Sure and no. So really we’re introducing a totally new problem which is totally different from what we’ve achieved earlier than. The brand new problem is the so-called good farming problem, which is opening our league to an entire new discipline of functions, as a result of we at the moment are agriculture. We’re engaged on this with Studica, a robotics firm from Canada, whose {hardware} we’re utilizing. We’ve already given it a strive on the German Open. This problem comes with some new specifics and one these is that the groups solely get the robotic shortly earlier than the competitors. So that they don’t actually know the robotic largely upfront, and they should assemble, program and design the robotic in a really quick time. To compensate for this, we cut back the quantity of optimization and robustness that’s vital. As a result of it’s an agriculture setting, we’ve totally different objects, like fruits, that the groups must deal with. This makes it a little bit bit extra sophisticated as a result of fruits have a extra arbitrary form and totally different ranges of ripeness that have to be detected. We even have some grapes which might be hanging on a wall, which is a totally totally different type of manipulation job than earlier than, as a result of earlier than the groups simply wanted to understand issues from surfaces, however now they really want to pluck stuff from a wall in a dependable means. So that is the brand new problem. It additionally comes with quite a lot of software program challenges as a result of the computational energy of this robotic could be very restricted as a result of it solely has a Raspberry Pi. For instance, there’s not quite a lot of picture processing attainable, particularly in comparison with the present robots, a few of which even have GPUs embedded.

Is there a selected a part of the {hardware} or software program that you simply’ve seen among the greatest developments in during the last yr or so?

Yeah, I believe one huge change I noticed over time was a swap from customized neural networks for object detection to off-the-shelf elements. So just about all groups these days use YOLO networks, which you will get pre-trained they usually simply deploy them on GPUs that they embedded into their robots. That is additionally one of many explanation why the robots actually grew in measurement over the previous couple of years as a result of they wanted house for the bigger computational energy. This really made it attainable for lots of groups to reliably detect the objects. Object detection was an enormous drawback at first of the league and these days it’s probably not an enormous situation – most groups are actually good at that. Generally they’re a little bit bit startled with decoy objects – these are objects within the enviornment that aren’t actually a part of the duty, and they’re unknown to the groups beforehand. Generally they’re, let’s say, evil decoys that appear like an object and there’s some mismatches that the groups do, however that is turning into very uncommon.

I believe the second huge change is a swap to bigger manipulators with extra levels of freedom. So at first, everybody had a really small manipulator with solely 5 levels of freedom, which restricted the working vary, and these days just about all groups have a six diploma of freedom manipulator with a wide range. Which means that they don’t want to maneuver their robotic when they’re in entrance of a workstation, which makes them a lot quicker and in addition rather more exact.

May you discuss in regards to the future plans for the League?

There are some things we’re enthusiastic about.

As regards to the competitors itself, we had a dialogue with the groups about what they’re concerned about doing sooner or later. Two issues got here up that they actually wish to have. One is cell obstacles, so they need different objects to maneuver autonomously via the world. We’re within the course of of making that, in cooperation with EduArt, which is an organization from Germany that additionally supplies small academic robots. And the second factor we wish to introduce is a type of humanoid robotic that the groups can use to truly deal with particular manipulation duties that can not be achieved merely with a robotic manipulator.

By way of creating an entry-level League, we’ve been engaged on this and one potential concept is to make use of the smart-farming problem because the entry level. Via the collaboration with Studica, we are able to present groups with the robotic they usually get to maintain it after the competitors. On the German Open I spoke to the Quickly-Manufactured Rescue League about this crossover between the Rescue and the Junior Leagues, and they’re very eager to collaborate.

We’re additionally speaking with the Logistics League. Their sponsor, Festo, has dropped out of their league and now they should reorganize. We’re questioning if it might be worthwhile to deliver our leagues nearer collectively, and even fuse them collectively to a single RoboCup Industrial league. The Logistics League desires to do extra manipulation, and the @Work League desires to do extra planning, so we’re closing the hole naturally between the 2. Nevertheless, that is only a thought for the time being – we have to see how the groups react to that.

About Christoph

Christoph Steup is an lively researcher specializing in varied fields of robotics, together with swarm robotics, precision farming, and weather-resilient autonomous driving. He presently works on the Fraunhofer Institute for Transportation and Infrastructure Techniques (IVI), the place he leads the Swarm Expertise Group. Previous to this function, he headed the Computational Intelligence in Robotics group on the Otto von Guericke College Magdeburg. Christoph’s involvement with RoboCup started in 2015 when he joined the robOTTO staff of Otto von Guericke College as staff chief. His contributions to the RoboCup group expanded as he turned a member of the Technical Committee for the @Work League in 2017. In 2019, he additional superior his participation by becoming a member of the Government Committee of the league.