An interview with Nicolai Ommer: the RoboCupSoccer Small Dimension League

Kick-off in a Small Dimension League match. Picture credit score: Nicolai Ommer.

RoboCup is a global scientific initiative with the objective of advancing the cutting-edge of clever robots, AI and automation. The annual RoboCup occasion is because of happen from 15-21 July in Salvador, Brazil. The Soccer element of RoboCup includes quite a few Leagues, with one in every of these being the Small Dimension League (SSL). We caught up with Government Committee member Nicolai Ommer to search out out extra concerning the SSL, how the auto referees work, and the way groups use AI.

May begin by giving us a fast introduction to the Small Dimension League?

Within the Small Dimension League (SSL) we’ve 11 robots per group – the one bodily RoboCup soccer league to have the total variety of gamers. The robots are small, cylindrical robots on wheels they usually can transfer in any route. They’re self-built by the groups, so groups need to do each the {hardware} and the programming, and a number of issues need to work collectively to make a group work. The AI is central. We don’t have brokers, so groups have a central pc on the area the place they will do all of the computation after which they ship the instructions to the robots in numerous abstractions. Some groups will simply ship velocity instructions, different groups ship a goal.

Now we have a central imaginative and prescient system – that is maintained by the League, and has been since 2010. There are cameras above the sphere to trace all of the robots and the ball, so everybody is aware of the place the robots are.

The robots can transfer as much as 4 meters per second (m/s), after this level it will get fairly unstable for the robots. They’ll change route in a short time, and the ball might be kicked at 6.5 m/s. It’s fairly quick and we’ve already needed to restrict the kick pace. Beforehand we had a restrict of 8 m/s and earlier than that 10m/s. Nonetheless, no robotic can catch a ball with this pace, so we determined to cut back it and put extra deal with passing. This offers the keeper and the defenders an opportunity to truly intercept a kick.

It’s so quick that for people it’s fairly obscure all of the issues which might be happening. And that’s why, some years in the past, we launched auto refs, which assist lots in monitoring, particularly issues like collisions and so forth, the place the human referee can’t watch all the pieces on the identical time.

How do the auto refs work then, and is there a couple of working on the identical time?

Once we developed the present system, to maintain issues truthful, we determined to have a number of implementations of an auto ref system. These unbiased methods implement the identical guidelines after which we do a majority vote on the selections.

To do that we wanted a center element, so some years in the past I began this mission to have a brand new recreation controller. That is the consumer interface (UI) for the human referee who sits at a pc. Within the UI you see the present recreation state, you may manipulate the sport state, and this element coordinates the auto refs. The auto refs can join and report fouls. If just one auto ref detects the foul, it received’t depend it. However, if each auto refs report the foul throughout the time window, then it’s counted. A part of the problem was to make this all visible for the operator to know. The human referee has the final phrase and makes the ultimate resolution.

We managed to determine two implementations. The goal was to have three implementations, which makes it simpler to kind a majority. Nonetheless, it nonetheless works with simply two implementations and we’ve had this for a number of years now. The implementations are from two totally different groups who’re nonetheless lively.

How do the auto refs cope with collisions?

We are able to detect collisions from the information. Nonetheless, even for human referees it’s fairly onerous to find out who was at fault when two robots collide. So we needed to simply outline a rule, and all of the implementations of the auto ref implement the identical rule. We wrote within the rulebook actually particularly the way you calculate if a collision occurred and who was at fault. The primary consideration relies on the rate – under 1.5m/s it’s not a collision, above 1.5m/s it’s. There’s additionally one other issue, referring to the angle calculation, that we additionally keep in mind to find out which robotic was at fault.

What else do the auto refs detect?

Different fouls embody the kick pace, after which there’s fouls referring to the adherence to regular recreation process. For instance, when the opposite group has a free kick, then the opposing robots ought to preserve a sure distance from the ball.

The auto refs additionally observe non-fouls, in different phrases recreation occasions. For instance, when the ball leaves the sphere. That’s the most typical occasion. This one is definitely not really easy to detect, notably if there’s a chip kick (the place the ball leaves the taking part in floor). With the digital camera lens, the parabola of the ball could make it seem like it’s exterior the sphere of play when it isn’t. You want a strong filter to cope with this.

Additionally, when the auto refs detect a objective, we don’t belief them utterly. When a objective is detected, we name it a “doable objective”. The match is halted instantly, all of the robots cease, and the human referee can verify all of the out there knowledge earlier than awarding the objective.

You’ve been concerned within the League for quite a few years. How has the League and the efficiency of the robots developed over that point?

My first RoboCup was in 2012. The introduction of the auto refs has made the play much more fluent. Earlier than this, we additionally launched the idea of ball placement, so the robots would place the ball themselves for a free kick, or kick off, for instance.

From the {hardware} facet, the principle enchancment in recent times has been dribbling the ball in one-on-one conditions. There has additionally been an enchancment within the specialised expertise carried out by robots with a ball. For instance, some years in the past, one group (ZJUNlict) developed robots that would pull the ball backwards with them, transfer round defenders after which shoot on the objective. This was an sudden motion, which we hadn’t seen earlier than. Earlier than this you needed to do a cross to trick the defenders. Our group, TIGERs Mannheim, has additionally improved on this space now. Nevertheless it’s actually tough to do that and requires a number of tuning. It actually is dependent upon the sphere, the carpet, which isn’t standardized. So there’s slightly little bit of luck that your particularly constructed {hardware} is definitely performing effectively on the competitors carpet.

The Small Dimension League Grand Remaining at RoboCup 2024 in Eindhoven, Netherlands. TIGERs Mannheim vs. ZJUNlict. Video credit score: TIGERs Mannheim. You’ll find the TIGERs’ YouTube channel right here.

What are a few of the challenges within the League?

One massive problem, and likewise perhaps it’s a great factor for the League, is that we’ve a number of undergraduate college students within the groups. These college students have a tendency to go away the groups after their Bachelor’s or Grasp’s diploma, the group members all change fairly commonly, and that signifies that it’s tough to retain data within the groups. It’s a problem to maintain the efficiency of the group; it’s even onerous to breed what earlier members achieved. That’s why we don’t have giant steps ahead, as a result of groups need to repeat the identical issues when new members be part of. Nonetheless, it’s good for the scholars as a result of they actually study lots from the expertise.

We’re repeatedly engaged on figuring out issues which we will make out there for everybody. In 2010 the imaginative and prescient system was established. It was an enormous issue, which means that groups didn’t need to do pc imaginative and prescient. And we’re presently establishing requirements for wi-fi communication – that is presently finished by everybody on their very own. We wish to advance the League, however on the identical time, we additionally wish to have this nature of having the ability to study, having the ability to do all of the issues themselves in the event that they wish to.

You actually need to have a group of individuals from totally different areas – mechanical engineering, electronics, mission administration. You additionally need to get sponsors, and it’s important to promote your mission, get college students in your group.

May you speak about a few of the AI parts to the League?

Most of our software program is script-based, however we apply machine studying for small, refined issues.

In my group, for instance, we do mannequin calibration with fairly easy algorithms. Now we have a selected mannequin for the chip kick, and one other for the robotic. The wheel friction is sort of sophisticated, so we give you a mannequin after which we gather the information and use machine studying to detect the parameters.

For the precise match technique, one good instance is from the group CMDragons. One 12 months you possibly can actually observe that that they had skilled their mannequin in order that, as soon as they scored objective, they upvoted the technique that they utilized earlier than that. You would actually see that the opponent reacted the identical approach on a regular basis. They had been capable of rating a number of targets, utilizing the identical technique repeatedly, as a result of they realized that if one technique labored, they might use it once more.

For our group, the TIGERs, our software program could be very a lot based mostly on calculating scores for the way good a cross is, how effectively can a cross be intercepted, and the way we will enhance the scenario with a specific cross. That is hard-coded typically, with some geometry-based calculations, however there may be additionally some fine-tuning. If we rating a objective then we monitor again and see the place the cross got here from and we give bonuses on a few of the rating calculations. It’s extra sophisticated than this, after all, however generally it’s what we attempt to do by studying through the recreation.

Folks usually ask why we don’t do extra with AI, and I believe the principle problem is that, in comparison with different use circumstances, we don’t have that a lot knowledge. It’s onerous to get the information. In our case we’ve actual {hardware} and we can’t simply do matches all day lengthy for days on finish – the robots would break, they usually have to be supervised. Throughout a contest, we solely have about 5 to seven matches in whole. In 2016, we began to file all of the video games with a machine-readable format. All of the positions are encoded, together with the referee selections, and all the pieces is in a log file which we publish centrally. I hope that with this rising quantity of information we will really apply some machine studying algorithms to see what earlier matches and former methods did, and perhaps get some insights.

What plans do you have got on your group, the TIGERs?

Now we have really received the competitors for the final 4 years. We hope that there will probably be another groups who can problem us. Our defence has probably not been challenged so we’ve a tough time discovering weaknesses. We really play towards ourselves in simulation.

One factor that we wish to enhance on is precision as a result of there may be nonetheless some handbook work to get all the pieces calibrated and dealing as exactly as we wish it. If some small element shouldn’t be working, for instance the dribbling, then it dangers the entire match. So we’re engaged on making all these calibration processes simpler, and to do extra computerized knowledge processing to find out the perfect parameters. In recent times we’ve labored lots on dribbling within the 1 vs 1 conditions. This has been a very massive enchancment for us and we’re nonetheless engaged on that.

About Nicolai

Nicolai Ommer is a Software program Engineer and Architect at QAware in Munich, specializing in designing and constructing strong software program methods. He holds a B.Sc. in Utilized Pc Science and an M.Sc. in Autonomous Techniques. Nicolai started his journey in robotics with Staff TIGERs Mannheim, taking part in his first RoboCup in 2012. His dedication led him to hitch the RoboCup Small Dimension League Technical Committee and, in 2023, the Government Committee. Keen about innovation and collaboration, Nicolai combines tutorial perception with sensible expertise to push the boundaries of clever methods and contribute to the worldwide robotics and software program engineering communities.