Tackling The 3D Simulation League: An Interview With Klaus Dorer And Stefan Glaser

A screenshot from the model new simulator that may be trialled for a specific downside at RoboCup2025.

The annual RoboCup event, the place teams acquire from all through the globe to take part in competitions all through various leagues, will this yr occur in Brazil, from 15-21 July. Upfront of kick-off, we spoke to 2 members of the RoboCup Soccer 3D Simulation League: Govt Committee Member Klaus Dorer, and Stefan Glaser, who’s on the Repairs Committee and who has been recently making a model new simulator for the League.

Would possibly start by merely giving us a quick introduction to the Simulation League?
Klaus Dorer: There are two Simulation Leagues in Soccer: the 2D Simulation League and the 3D Simulation League. The 2D Simulation League, as a result of the determine suggests, is a flat league the place the avid gamers and ball are simulated with simplified physics and the first focus is on group approach. The 3D Simulation League is means nearer to precise robots; it simulates 11 versus 11 Nao robots. The extent of administration is like with precise robots, the place you progress each motor of the legs and the arms and so forth to achieve movement.

I understand that you simply’ve been engaged on a model new simulator for the 3D League. What was the thought behind this new simulator?
Klaus: The intention is to hold us nearer to the {{hardware}} leagues so that the simulator could also be further useful. The current simulator that we use inside the 3D Simulation League is called SimSpark. It was created inside the early 2000s with the intention of making it attainable to play 11 vs 11 avid gamers. With the {{hardware}} constraints of that time, there wanted to be some compromises on the physics to have the power to simulate 22 avid gamers on the same time. So the simulation is bodily significantly actual trying, nonetheless not inside the sense that it’s easy to transpose it to an precise Nao robotic.

Stefan Glaser: The thought for making a model new simulator has been spherical for plenty of years. SimSpark is a very extremely efficient simulation framework. The underside framework is space neutral (not soccer explicit) and explicit simulations are realized by the use of plugins. It helps plenty of physics engines inside the backend and provides a flexible scripting interface for configuration and variations of the simulation. Nonetheless, all this flexibility comes with the value of complexity. Together with that, SimSpark makes use of custom-made robotic model specs and communication protocols, limiting the amount of accessible robotic fashions and requiring teams to develop custom-made communication layers only for talking with SimSpark. On account of this, SimSpark has not been extensively adopted inside the RoboCup group.

With the model new simulator, I need to deal with these two details: complexity and standardization. Throughout the ML group, the MuJoCo physics engine has flip into a really modern various for learning environments after Google DeepMind acquired it and launched it open provide. Its necessities for world and robotic model specs are extensively adopted regionally and there exist a complete lot of ready-to-use robotic model specs for every kind of digital along with real-world robots. In the midst of last yr, they (MuJoCo) added a attribute which helps you to manipulate the world illustration all through simulation (together with and eradicating objects to / from the simulation whereas preserving the simulation state). That’s one necessary requirement we’ve acquired inside the simulation league, the place we start with an empty topic after which the brokers be part of on demand and sort the teams. When this attribute has been added, I decided to make a step forward and try and implement a model new simulator for the 3D Simulation League based on MuJoCo. Initially, I wanted to start progress in C/C++ to achieve most effectivity, nonetheless then decided to start in Python to cut back complexity and make it further accessible for various builders. I started progress on Easter Monday so it’s not even three months earlier!

I really feel it’s maybe useful to make clear a bit bit further regarding the setup of our league and the requirements of the simulator. If we take the FIFA sport (in your favorite gaming machine) for instance, there could also be one simulation occurring which simulates 22 avid gamers and the selection making is part of the simulation having full entry to the state of the world. Throughout the 3D Simulation League we’ve acquired two teams with 11 robots on the sector, nonetheless we even have 22 explicit individual agent softwares which might be linked to the simulation server, each controlling one single robotic. Each linked agent solely receives sensor information related to their robotic inside the simulation. They’re moreover solely allowed to talk by the use of the server – there isn’t a direct communication between the brokers allowed in Simulation League. So we’ve acquired a traditional setup the place the simulation server has to have the power to accept as a lot as 22 connections and deal with the state of affairs there. This efficiency has been the primary focus for me for the ultimate couple of months and this half is already working correctly. Teams can be part of their brokers, which is ready to get hold of sensor information and may actuate joints of the robotic inside the simulation and so forth. They’re moreover able to select completely completely different robotic fashions within the occasion that they like.

An illustration of the simulator set-up.

Presumably the model new simulator has a higher illustration of the physics of an precise robotic.
Klaus: Exactly. For example, how the motors are managed is now a bit completely completely different and much nearer to precise robots. So as soon as I did my first experiments, I seen the robotic collapse and I assumed it was exactly how an precise robotic would collapse! In SimSpark we moreover had falling robots nonetheless the motor administration inside the new simulator is completely completely different. Now you’ll be capable of administration the motors by velocity, by strain, by place, which is far more versatile – it’s nearer to what everyone knows from precise robots.

I really feel that, on the very least initially, it’s going to doubtless be more durable for the Simulation League teams to get the robots to do what they want them to do, on account of it’s further actual trying. For example, in SimSpark the underside contact was far more forgiving. So for those who occur to step laborious on the underside, you don’t fall immediately with a SimSpark robotic nonetheless with a MuJoCo robotic this may be far more actual trying. Actually, in precise robots flooring contact is significantly a lot much less forgiving.

I had a question regarding the imaginative and prescient aspect – how do the individual brokers “see” the place of the other brokers on the sector?
Stefan: We simulate a digital imaginative and prescient pipeline on the server side. You’ll have a restricted topic of view of ±60° horizontally and vertically. Inside that topic of view you’ll detect the top, the arms, the toes of various avid gamers, or the ball, as an illustration, or completely completely different choices of the sector. Similar to frequent real-world imaginative and prescient pipelines, each detection consists of a label, a course vector and the house information. The info has some noise on it like precise robots have, too, nonetheless teams don’t should course of digicam pictures. They get the detections straight from the simulation server.

We’ve beforehand had a dialogue about transferring within the path of getting digicam pictures of the simulation to mix into the imaginative and prescient pipeline on the agent side. This was not at all truly actual trying in SimSpark with the implementation we had there. Nonetheless, it must be attainable with MuJoCo. Nonetheless, for the first mannequin, I used the an identical method the traditional simulator dealt with the imaginative and prescient. Due to this teams don’t need to coach a imaginative and prescient model, and don’t should cope with digicam pictures to get started. This reduces the load significantly and as well as shifts the primary focus of the problem within the path of motion and selection making.

Will the simulator be used at RoboCup 2025?
Stefan: We plan to have an issue with a model new simulator and I’ll try and current some demo video video games. In the intervening time it’s not going in a state the place you’ll be capable of play an entire opponents.

Klaus: That’s usually how we proceed with new simulators. We’d not switch from one to the other with none intermediate step. We might have an issue this yr at RoboCup 2025 with the model new MuJoCo simulator the place each collaborating group will try to teach the robotic to kick as far as attainable. So, we gained’t be collaborating in an entire sport, we acquired’t have plenty of robots, solely a single robotic stepping in entrance of the ball and kicking the ball. That’s the technical downside for this yr. Teams will get an considered how the simulator works, and we’ll get an considered what must be modified inside the simulator to proceed.

This new downside may be voluntary, so we’re not optimistic what variety of teams will participate. Our group (MagmaOffenburg) will certainly take part. Will in all probability be attention-grabbing to see how correctly the teams perform on account of no person is conscious of how far an outstanding kick is on this simulator. It’s a bit like in Elements One when the ideas change and no person is conscious of which group can be the primary group.

Do you’ve gotten an considered how quite a bit adaptation teams ought to make if and everytime you switch to the model new simulator for the entire matches?
Stefan: As a long-term member of 3D Simulation League, I do know the earlier simulator SimSpark pretty correctly, and know the protocols involved and the way in which the processes work. So the first mannequin of the model new simulator is designed to utilize the an identical basic protocol, the an identical sensor information, and so forth. The thought is that the teams can use the model new simulator with minimal effort in adapting their current agent software program program. In order that they have to be succesful to get started pretty fast.

Although, when designing a model new platform, I need to take the prospect to make a step forward by the use of protocols, on account of I moreover want to mix completely different Leagues inside the long-term. They usually produce different administration mechanisms, and they also don’t use the an identical protocol that’s excellent in 3D Simulation. Subsequently there must be some flexibility eventually. Nevertheless for the first mannequin, the thought was to get the Simulation League ready with minimal effort.

Klaus: The large thought is that this isn’t merely used inside the 3D Simulation league, however as well as as a useful simulator for the Humanoid League and as well as for the Commonplace Platform League (SPL). So if that appears to be true, then it’s going to doubtless be completely worthwhile. For the Kick Drawback this yr, as an illustration, we use a T1 robotic that could be a Humanoid League robotic.

Would possibly you say one factor about this simulation to precise world (Sim2Real) aspect?
Stefan: We’d desire it to be attainable for the motions and behaviors inside the simulator to be ported to precise robots. From my viewpoint, it may very well be useful the other method spherical too.

We, as a Simulation League, usually develop for the Simulation League and as a result of this reality need to get the behaviors working on an precise robotic. Nevertheless the {{hardware}} teams usually have an an identical topic after they want to test high-level alternative making. They may have two to five robots on the sector, and within the occasion that they want to play a high-level decision-making match and put together in that regard, they always must deploy a complete lot of robots. If moreover they want to have an opponent, they should double the amount of robots in order to play a sport to see how the approach would find yourself. The Sim2Real aspect can be attention-grabbing for these teams, on account of they have to be succesful to take what they deployed on the precise robotic and it additionally must work inside the simulation. They’ll then use the simulation to educate high-level skills like group play, participant positioning and so forth, which is a troublesome aspect for the precise robotic leagues like SPL or the Humanoid Leagues.

Klaus: And the reason everyone knows it’s as a result of we’ve acquired a bunch inside the Simulation League and we’ve acquired a bunch inside the Humanoid League. So that’s another reason why we’re keen to hold these things nearer collectively.

How does the refereeing work inside the Simulation League?
Klaus: A pleasing issue about Simulation Leagues is that there’s a program which is conscious of the particular state of the world so we’re capable of assemble inside the referee contained within the simulator and it’ll not fail. For points like offside, whether or not or not the ball handed the target line, that’s fail safe. All the referee decisions are taken by the system itself. We now have a human referee nonetheless they not at all should intervene. Nonetheless, there are situations the place we wish artificial intelligence to play a job. This isn’t presently the case in SimSpark on account of the ideas are all laborious coded. We now have a complete lot of fouls that are debatable. For example, there are numerous fouls that teams agree mustn’t have been a foul, and completely different fouls that aren’t referred to as that must have been. It may very well be a nice AI learning job to get some situations judged by human referees after which put together an AI model to raised resolve the ideas for what’s a foul and what isn’t a foul. Nevertheless that’s presently not the case.

Stefan: On the model new simulator I’m not that far into the occasion that I’ve carried out the automated referee however. I’ve some basic algorithm which progress the game as such, nonetheless judging fouls and deciding on explicit situations won’t be however carried out inside the new simulator.

What are the next steps for creating the simulator?
Stefan: One in all many subsequent major steps may be to refine the physics simulation. For example, regardless that there exists a ball inside the simulation, it’s not however very nicely refined. There are a complete lot of physics parameters which we’ve acquired to find out on to reflect the precise world just about nearly as good as attainable. It is going to doable require a sequence of experiments in order to get to the right values for quite a few aspects. On this aspect I’m hoping for some engagement of the group, because it’s a pleasant evaluation various and I personally would favor the group to find out on a usually accepted parameter set based on a level of proof that I can’t merely current all alone. So in case any person is interested in refining the physics of the simulation such that it most interesting shows the precise world, you may be welcome to hitch!

One different major subsequent step can be the expansion of the automated referee of the soccer simulation, deciding on fouls, coping with misbehaving brokers and so forth. Throughout the first mannequin, foul conditions will doable be judged by an educated system notably designed for this aim. The simulation league has developed a set of foul state of affairs specs which I plan to adapt. In a second step, I need to mix and help the occasion of AI based foul detection fashions. Nevertheless yeah, one step after the other.

What are you notably attempting forward to at RoboCup2025?
Klaus: Correctly, with our group we’ve acquired been vice world champion seven situations in a row. This yr we’re truly hoping to make it to world champion. We’re very expert in getting losses in finals and this yr we’re attempting forward to altering that, from a bunch perspective.

Stefan: I’m going to Brazil in order to advertise the simulator, not just for the Simulation League, however as well as all through the boundaries for the Humanoid Leagues and the SPL Leagues. I really feel that this simulator is an outstanding probability to hold of us from all the leagues collectively. I’m notably inside the explicit requirements of all the teams of the completely completely different leagues. This understanding will help me tailor the model new simulator within the path of their desires. That’s definitely considered one of my major highlights for this yr, I’d say.

You’ll uncover out further regarding the new simulator on the endeavor webpage, and from the documentation.

Klaus Dorer is professor for artificial intelligence, autonomous applications and software program program engineering at Offenburg School, Germany. He’s moreover a member of the Institute for Machine Learning and Analytics IMLA. He has been group chief of the RoboCup simulation league teams magmaFreiburg (since 1999), dwelling applications, magmaFurtwangen and is now group chief of magmaOffenburg since 2009. Since 2014, he has moreover been part of the humanoid grownup measurement league group Sweaty.

Stefan Glaser is instructing assistant for artificial intelligence and intelligent autonomous applications on the Offenburg School, Germany. He has been part of the RoboCup simulation league group magmaOffenburg since 2009 and the RoboCup humanoid grownup measurement league group Sweaty since 2014.