Tackling the 3D Simulation League: an interview with Klaus Dorer and Stefan Glaser

A screenshot from the brand new simulator that will likely be trialled for a particular problem at RoboCup2025.

The annual RoboCup occasion, the place groups collect from throughout the globe to participate in competitions throughout quite a lot of leagues, will this yr happen 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 Upkeep Committee and who has been not too long ago growing a brand new simulator for the League.

May begin by simply giving us a fast 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, because the title suggests, is a flat league the place the gamers and ball are simulated with simplified physics and the principle focus is on crew technique. The 3D Simulation League is way nearer to actual robots; it simulates 11 versus 11 Nao robots. The extent of management is like with actual robots, the place you progress every motor of the legs and the arms and so forth to realize motion.

I perceive that you’ve been engaged on a brand new simulator for the 3D League. What was the thought behind this new simulator?
Klaus: The intention is to deliver us nearer to the {hardware} leagues in order that the simulator might be extra helpful. The present simulator that we use within the 3D Simulation League is named SimSpark. It was created within the early 2000s with the intention of constructing it doable to play 11 vs 11 gamers. With the {hardware} constraints of that point, there needed to be some compromises on the physics to have the ability to simulate 22 gamers on the identical time. So the simulation is bodily considerably life like, however not within the sense that it’s simple to transpose it to an actual Nao robotic.

Stefan Glaser: The thought for growing a brand new simulator has been round for a couple of years. SimSpark is a really highly effective simulation framework. The bottom framework is area impartial (not soccer particular) and particular simulations are realized by way of plugins. It helps a number of physics engines within the backend and gives a versatile scripting interface for configuration and variations of the simulation. Nevertheless, all this flexibility comes with the worth of complexity. Along with that, SimSpark makes use of customized robotic mannequin specs and communication protocols, limiting the quantity of accessible robotic fashions and requiring groups to develop customized communication layers just for speaking with SimSpark. On account of this, SimSpark has not been broadly adopted within the RoboCup neighborhood.

With the brand new simulator, I want to handle these two main points: complexity and standardization. Within the ML neighborhood, the MuJoCo physics engine has develop into a highly regarded alternative for studying environments after Google DeepMind acquired it and launched it open supply. Its requirements for world and robotic mannequin specs are broadly adopted in the neighborhood and there exist a whole lot of ready-to-use robotic mannequin specs for all kinds of digital in addition to real-world robots. In the course of final yr, they (MuJoCo) added a function which lets you manipulate the world illustration throughout simulation (including and eradicating objects to / from the simulation whereas preserving the simulation state). That is one important requirement we’ve got within the simulation league, the place we begin with an empty discipline after which the brokers join on demand and kind the groups. When this function has been added, I made a decision to make a step ahead and attempt to implement a brand new simulator for the 3D Simulation League primarily based on MuJoCo. Initially, I needed to start out growth in C/C++ to realize most efficiency, however then determined to start out in Python to cut back complexity and make it extra accessible for different builders. I began growth on Easter Monday so it’s not even three months outdated!

I believe it could be helpful to clarify a bit of bit extra concerning the setup of our league and the necessities of the simulator. If we take the FIFA recreation (in your favourite gaming gadget) for instance, there may be one simulation taking place which simulates 22 gamers and the choice making is a part of the simulation having full entry to the state of the world. Within the 3D Simulation League we’ve got two groups with 11 robots on the sphere, however we even have 22 particular person agent softwares that are linked to the simulation server, every controlling one single robotic. Every linked agent solely receives sensor data associated to their robotic within the simulation. They’re additionally solely allowed to speak by way of the server – there isn’t any direct communication between the brokers allowed in Simulation League. So we’ve got a common setup the place the simulation server has to have the ability to settle for as much as 22 connections and handle the scenario there. This performance has been the main focus for me for the final couple of months and this half is already working properly. Groups can join their brokers, which is able to obtain sensor data and may actuate joints of the robotic within the simulation and so forth. They’re additionally capable of choose completely different robotic fashions in the event that they like.

An illustration of the simulator set-up.

Presumably the brand new simulator has a greater illustration of the physics of an actual robotic.
Klaus: Precisely. For instance, how the motors are managed is now a bit completely different and far nearer to actual robots. So once I did my first experiments, I noticed the robotic collapse and I believed it was precisely how an actual robotic would collapse! In SimSpark we additionally had falling robots however the motor management within the new simulator is completely different. Now you may management the motors by velocity, by drive, by place, which is rather more versatile – it’s nearer to what we all know from actual robots.

I believe that, a minimum of initially, it will likely be harder for the Simulation League groups to get the robots to do what they need them to do, as a result of it’s extra life like. For instance, in SimSpark the bottom contact was rather more forgiving. So in the event you step arduous on the bottom, you don’t fall instantly with a SimSpark robotic however with a MuJoCo robotic this will likely be rather more life like. Certainly, in actual robots floor contact is considerably much less forgiving.

I had a query concerning the imaginative and prescient facet – how do the person brokers “see” the place of the opposite brokers on the sphere?
Stefan: We simulate a digital imaginative and prescient pipeline on the server aspect. You will have a restricted discipline of view of ±60° horizontally and vertically. Inside that discipline of view you’ll detect the top, the arms, the ft of different gamers, or the ball, for instance, or completely different options of the sphere. Much like frequent real-world imaginative and prescient pipelines, every detection consists of a label, a route vector and the gap data. The knowledge has some noise on it like actual robots have, too, however groups don’t have to course of digital camera pictures. They get the detections straight from the simulation server.

We’ve beforehand had a dialogue about shifting in the direction of getting digital camera pictures of the simulation to combine into the imaginative and prescient pipeline on the agent aspect. This was by no means actually life like in SimSpark with the implementation we had there. Nevertheless, it needs to be doable with MuJoCo. Nevertheless, for the primary model, I used the identical method the standard simulator handled the imaginative and prescient. Because of this groups don’t want to coach a imaginative and prescient mannequin, and don’t have to deal with digital camera pictures to get began. This reduces the load considerably and likewise shifts the main focus of the issue in the direction of movement and resolution making.

Will the simulator be used at RoboCup 2025?
Stefan: We plan to have a problem with a brand new simulator and I’ll attempt to present some demo video games. In the intervening time it’s probably not in a state the place you may play a complete competitors.

Klaus: That’s normally how we proceed with new simulators. We might not transfer from one to the opposite with none intermediate step. We can have a problem this yr at RoboCup 2025 with the brand new MuJoCo simulator the place every collaborating crew will attempt to train the robotic to kick so far as doable. So, we is not going to be taking part in a complete recreation, we received’t have a number of robots, only a single robotic stepping in entrance of the ball and kicking the ball. That’s the technical problem for this yr. Groups will get an thought of how the simulator works, and we’ll get an thought of what must be modified within the simulator to proceed.

This new problem will likely be voluntary, so we’re not certain what number of groups will take part. Our crew (MagmaOffenburg) will definitely participate. It is going to be attention-grabbing to see how properly the groups carry out as a result of nobody is aware of how far a great kick is on this simulator. It’s a bit like in System One when the foundations change and nobody is aware of which crew would be the main crew.

Do you’ve an thought of how a lot adaptation groups should make if and once you transfer to the brand new simulator for the complete matches?
Stefan: As a long-term member of 3D Simulation League, I do know the outdated simulator SimSpark fairly properly, and know the protocols concerned and the way the processes work. So the primary model of the brand new simulator is designed to make use of the identical primary protocol, the identical sensor data, and so forth. The thought is that the groups can use the brand new simulator with minimal effort in adapting their present agent software program. So they need to have the ability to get began fairly quick.

Though, when designing a brand new platform, I want to take the chance to make a step ahead when it comes to protocols, as a result of I additionally need to combine different Leagues within the long-term. They normally produce other management mechanisms, and so they don’t use the identical protocol that’s distinguished in 3D Simulation. Subsequently there must be some flexibility sooner or later. However for the primary model, the thought was to get the Simulation League prepared with minimal effort.

Klaus: The massive thought is that this isn’t simply used within the 3D Simulation league, but in addition as a helpful simulator for the Humanoid League and likewise for the Normal Platform League (SPL). So if that seems to be true, then it will likely be fully profitable. For the Kick Problem this yr, for instance, we use a T1 robotic that could be a Humanoid League robotic.

May you say one thing about this simulation to actual world (Sim2Real) facet?
Stefan: We’d prefer it to be doable for the motions and behaviors within the simulator to be ported to actual robots. From my viewpoint, it might be helpful the opposite method spherical too.

We, as a Simulation League, normally develop for the Simulation League and due to this fact want to get the behaviors working on an actual robotic. However the {hardware} groups normally have the same situation once they need to take a look at high-level resolution making. They may have two to 5 robots on the sphere, and in the event that they need to play a high-level decision-making match and practice in that regard, they all the time need to deploy a whole lot of robots. If additionally they need to have an opponent, they need to double the quantity of robots with the intention to play a recreation to see how the technique would prove. The Sim2Real facet can be attention-grabbing for these groups, as a result of they need to have the ability to take what they deployed on the actual robotic and it also needs to work within the simulation. They’ll then use the simulation to coach high-level abilities like crew play, participant positioning and so forth, which is a difficult facet for the actual robotic leagues like SPL or the Humanoid Leagues.

Klaus: And the rationale we all know it is because we’ve got a crew within the Simulation League and we’ve got a crew within the Humanoid League. In order that’s another excuse why we’re eager to deliver these items nearer collectively.

How does the refereeing work within the Simulation League?
Klaus: A pleasant factor about Simulation Leagues is that there’s a program which is aware of the actual state of the world so we will construct within the referee contained in the simulator and it’ll not fail. For issues like offside, whether or not the ball handed the aim line, that’s fail protected. All of the referee selections are taken by the system itself. We now have a human referee however they by no means have to intervene. Nevertheless, there are conditions the place we wish synthetic intelligence to play a job. This isn’t at the moment the case in SimSpark as a result of the foundations are all arduous coded. We now have a whole lot of fouls which can be debatable. For instance, there are numerous fouls that groups agree mustn’t have been a foul, and different fouls that aren’t referred to as that ought to have been. It will be a pleasant AI studying process to get some conditions judged by human referees after which practice an AI mannequin to raised decide the foundations for what’s a foul and what isn’t a foul. However that is at the moment not the case.

Stefan: On the brand new simulator I’m not that far into the event that I’ve carried out the automated referee but. I’ve some primary algorithm which progress the sport as such, however judging fouls and deciding on particular conditions isn’t but carried out within the new simulator.

What are the subsequent steps for growing the simulator?
Stefan: One of many subsequent main steps will likely be to refine the physics simulation. For example, regardless that there exists a ball within the simulation, it isn’t but very well refined. There are a whole lot of physics parameters which we’ve got to determine on to mirror the actual world pretty much as good as doable. This may seemingly require a sequence of experiments with the intention to get to the proper values for varied facets. On this facet I’m hoping for some engagement of the neighborhood, as it’s a nice analysis alternative and I personally would like the neighborhood to determine on a generally accepted parameter set primarily based on a degree of proof that I can’t simply present all on my own. So in case somebody is interested by refining the physics of the simulation such that it finest displays the actual world, you might be welcome to hitch!

One other main subsequent step would be the growth of the automated referee of the soccer simulation, deciding on fouls, dealing with misbehaving brokers and so forth. Within the first model, foul situations will seemingly be judged by an professional system particularly designed for this goal. The simulation league has developed a set of foul situation specs which I plan to adapt. In a second step, I want to combine and assist the event of AI primarily based foul detection fashions. However yeah, one step after the opposite.

What are you notably trying ahead to at RoboCup2025?
Klaus: Effectively, with our crew we’ve got been vice world champion seven instances in a row. This yr we’re actually hoping to make it to world champion. We’re very skilled in getting losses in finals and this yr we’re trying ahead to altering that, from a crew perspective.

Stefan: I’m going to Brazil with the intention to promote the simulator, not only for the Simulation League, but in addition throughout the boundaries for the Humanoid Leagues and the SPL Leagues. I believe that this simulator is a good probability to deliver individuals from all of the leagues collectively. I’m notably within the particular necessities of all of the groups of the completely different leagues. This understanding will assist me tailor the brand new simulator in the direction of their wants. That is one in every of my main highlights for this yr, I’d say.

You’ll find out extra concerning the new simulator on the challenge webpage, and from the documentation.

Klaus Dorer is professor for synthetic intelligence, autonomous methods and software program engineering at Offenburg College, Germany. He’s additionally a member of the Institute for Machine Studying and Analytics IMLA. He has been crew chief of the RoboCup simulation league groups magmaFreiburg (since 1999), dwelling methods, magmaFurtwangen and is now crew chief of magmaOffenburg since 2009. Since 2014, he has additionally been a part of the humanoid grownup dimension league crew Sweaty.

Stefan Glaser is instructing assistant for synthetic intelligence and clever autonomous methods on the Offenburg College, Germany. He has been a part of the RoboCup simulation league crew magmaOffenburg since 2009 and the RoboCup humanoid grownup dimension league crew Sweaty since 2014.