Self-supervised learning for soccer ball detection and beyond: interview with winners of the RoboCup 2025 best paper award

Discussion of the very best paper honor at the RoboCup 2025 seminar.

A crucial facet of self-governing soccer-playing robotics issues precise discovery of the round. This is the emphasis of job by Can Lin, Daniele Affinita, Marco Zimmatore, Daniele Nardi, Domenico Bloisi, and Vincenzo Suriani, which won the very best paper honor at the current RoboCupsymposium The seminar happens along with the yearly RoboCup competition, which this year was kept in Salvador, Brazil. We overtook several of the writers to discover even more regarding the job, just how their technique can be moved to applications past RoboCup, and their future prepare for the competitors.

Could you begin by providing us a short summary of the issue that you were attempting to address in your paper “Self-supervised Attribute Removal for Improved Sphere Discovery on Football Robots”?

Daniele Affinita: The major obstacle we dealt with was that deep knowing typically calls for a big quantity of identified information. This is not a significant issue for usual jobs that have actually currently been examined, since you can generally locate labeled datasets online. However when the job is very particular, like in RoboCup, you require to accumulate and classify the information on your own. That implies collecting the information and by hand annotating it prior to you can also begin using deep knowing. This procedure is not scalable and requires a considerable human initiative.

The concept behind our paper was to decrease this human initiative. We came close to the issue with self-supervised knowing, which intends to discover helpful depictions of the information. Nevertheless, deep knowing is basically regarding discovering concealed depictions from the readily available information.

Could you inform us a little bit much more regarding your self-supervised knowing structure and just how you tackled establishing it?

Daniele: To Start With, allow me present what self-supervised knowing is. It is a means of discovering the framework of the information without having accessibility to tags. This is generally done with what we call pretense jobs These are jobs that do not call for specific tags, yet rather manipulate the framework of the information. For instance, in our instance we dealt with pictures. You can arbitrarily mask some spots and educate the design to anticipate the missing out on components. By doing so, the design is required to discover purposeful attributes from the information.

In our paper, we enhanced the information by utilizing not just raw pictures yet additionally outside assistance. This originated from a bigger design which we describe as the educator This design was educated on a various job which is much more basic than the target job we went for. In this manner the bigger design can give assistance (an exterior signal) that aids the self-supervision to concentrate much more on the particular job we appreciate.

In our instance, we wished to anticipate a limited circle the round. To direct this, we utilized an exterior pretrained design (YOLO) for things discovery, which rather forecasts a loosened bounding box around the round. We can probably state that the bounding box, a rectangular shape, is much more basic than a circle. So in this feeling, we were attempting to make use of outside assistance that does not address specifically the underlying job.

Review of the information prep work pipe.

Were you able to evaluate this design out at RoboCup 2025?

Daniele: Yes, we released it at RoboCup 2025 and revealed fantastic enhancements over our previous standard, which was the design we utilized in 2024. Specifically, we observed that the last training calls for a lot less information. The design was additionally much more durable under various illumination problems. The problem we had with previous designs was that they were customized for particular scenarios. However naturally, all the places are various, the illumination and the illumination are various, there may be darkness on the area. So it’s truly vital to have a reputable design and we truly observed a fantastic renovation this year.

What’s your group name, and could you chat a little bit regarding the competitors and just how it went?

Daniele: So our group is SPQR. We are from Rome, and we have actually been completing in RoboCup for a long period of time.

Domenico Blois: We began in 1998, so we are among the earliest groups in RoboCup.

Daniele: Yeah, I had not been also birthed after that! Our group began with the four-legged robotics. And after that the organization changed much more in the direction of biped robotics since they are much more tough, they call for equilibrium and, total it’s tougher to stroll on simply 2 legs.

Our group has actually expanded a great deal throughout current years. We have actually been adhering to an extremely favorable fad, going from 9th location in 2019 to 3rd location at the German Open in 2025, and we obtained fourth location at RoboCup 2025. Our current success has actually brought in even more pupils to the group. So it’s sort of a loophole– you win much more, you bring in much more pupils, and you can function much more on the difficulties recommended by RoboCup.

SPQR group.

Domenico: I wish to include that additionally, from a study viewpoint, we have actually won 3 ideal paper honors in the last 5 years, and we have actually been suggesting some brand-new fads in the direction of, for instance, using LLMs for coding (as a robotic’s behavior generator under the guidance of a human instructor). So we are attempting to maintain the open research study area energetic in our group. We wish to win the suits yet we additionally wish to address the research study issues that are bound along with the competitors.

Among the vital payments of our paper is in the direction of using our formulas outside RoboCup. For instance, we are attempting to use the round detector in accuracy farming. We wish to make use of the exact same strategy to find spherical fruits. This is something that is truly vital for us; to leave the context of Robocup and to make use of Robocup devices for brand-new techniques in various other areas. So if we shed a suit, it’s not a huge bargain for us. We desire our pupils, our employee, to be open minded in the direction of using RoboCup as a beginning factor for comprehending synergy and for comprehending just how to handle rigorous target dates. This is something that RoboCup can offer us. We attempt to have a group that awaits every sort of obstacle, not just within RoboCup, yet additionally various other sorts of AI applications. Winning is not every little thing for us. We would certainly like to utilize our very own code and not win, than win making use of code established by others. This is not ideal for attaining top place, yet we wish to educate our pupils to be gotten ready for the research study that is beyond RoboCup.

You claimed that you have actually formerly won 2 various other ideal paper honors. What did those documents cover?

Domenico: So the last 2 ideal documents were sort of visionary documents. In one paper, we wished to offer an understanding in just how to make use of the viewers to aid the robotics rack up. For instance, if you support louder, the robotics have a tendency to kick the round. So this is something that is not really utilized in the competitors currently, yet is something much more in the direction of the 2050 obstacle. So we wish to picture just how it will certainly be ten years from currently.

The other paper was called “play everywhere“, so you can, for instance, have fun with various sorts of round, you can play outside, you can also play without a certain objective, you can play making use of Coca-Cola canisters as goalposts. So the robotic needs to have a basic strategy that is not connected to the particular area utilized in RoboCup. This remains in comparison to various other groups that are really particular. We have a various strategy and this is something that makes it harder for us to win the competitors. Nevertheless, we do not wish to win the competitors, we wish to attain this objective of having, in 2050, this suit in between the RoboCup champions and the FIFA Globe Mug champions.

I have an interest in what you claimed regarding moving the technique for round discovery to farming and various other applications. Could you state much more regarding that research study?

Vincenzo Suriani: Our laboratory has actually been associated with some various jobs associating with farming applications. The Embellishment job ranged from 2015– 2018. Extra lately, the CANOPIES project has actually concentrated on accuracy farming for long-term plants where farmworkers can effectively interact with groups of robotics to do agronomic treatments, like gathering or trimming.

We have an additional job that has to do with spotting and gathering grapes. There is a significant initiative in bringing understanding back from RoboCup to various other jobs, and the other way around.

Domenico: Our vision currently is to concentrate on the brand-new generation of humanoid robotics. We took part in a brand-new occasion, the Globe Humanoid Robotic Gamings, kept in Beijing in August 2025, since we wish to make use of the system of RoboCup for various other type of applications. The concept is to have a solitary system with software program that is stemmed from RoboCup code that can be utilized for various other applications. If you have a humanoid robotic that requires to relocate, you can recycle the exact same code from RoboCup since you can make use of the exact same stablizing, the exact same vision core, the exact same structure (essentially), and you can simply transform some components and you can have a totally various sort of application with the exact same robotic with essentially the exact same code. We wish to go in the direction of this concept of recycling code and having RoboCup as an examination bed. It is an extremely challenging examination bed, yet you can make use of the lead to various other areas and in various other applications.

Looking especially at RoboCup, what are your future prepare for the group? There are some huge adjustments prepared for the RoboCup Leagues, so could you additionally state just how this might influence your strategies?

Domenico: We have an extremely solid group and several of the employee will certainly do a PhD in the coming years. Among our targets was to maintain the pupils inside the college and the research study ward, and we achieved success in this, since currently they are really enthusiastic regarding the RoboCup competitors and regarding AI as a whole.

In regards to the adjustments, there will certainly be a brand-new organization within RoboCup that is a merging of the conventional system organization (SPL) and the humanoid kid-size organization. The humanoid adult-size organization will certainly continue to be, so we require to choose whether to sign up with the brand-new joined organization, or relocate to adult-sized robotics. Presently we do not have way too many information, yet what we understand is that we will certainly go in the direction of a brand-new age of robotics. We got robotics from Booster and we are currently obtaining an additional G1 robotic from Unitree. So we are attempting to have a full family members of brand-new robotics. And after that I assume we will certainly go in the direction of the organization that is picked by the various other groups in the SPL organization. But also for currently we are attempting to arrange an occasion in October in Rome with 2 various other groups to trade concepts and to recognize where we wish to go. There will certainly additionally be a workshop to review the research study side.

Vincenzo: We are additionally in conversation regarding the very best dimension of robotic for the competitors. We are mosting likely to have 2 various settings, since robotics are coming to be less costly and there are groups that are pressing to relocate quicker to a larger system. On the various other hand, there are groups that wish to stick to a smaller sized system in order to research on multi representatives. We have actually seen a great deal of applications for a solitary robotic yet very few applications with a collection of robotics that are complying. And this has actually been traditionally among the core components of research study we performed in RoboCup, and additionally beyond RoboCup.

There are a lot of perspectives on which robotic dimension to make use of, since there are a number of elements, and we do not recognize just how rapid the globe will certainly transform in 2 or 3 years. We are attempting to form the regulations and the problems to bet following year, yet, due to just how rapidly points are transforming, we do not recognize what the very best choice will certainly be. And additionally the research study we are mosting likely to do will certainly be influenced by the choice we make on this.

There will certainly be some adjustments to various other organizations in the future also; the little and center dimensions will certainly enclose 2 years possibly, and the simulation organization additionally. A great deal will certainly occur in the following 5 years, possibly greater than throughout the last 10-15 years. This is an essential year since the choices are based upon what we can see, what we can identify in the future, yet we do not have all the info we require, so it will certainly be tough.

For instance, the SPL has a huge, possibly the most significant, area amongst the RoboCup organizations. We have a great deal of groups that are organizing by rate of interest therefore there are groups that are adhering to working with this particular issue with a certain system and groups that are attempting to relocate to an additional system and an additional issue. So also inside the exact same area we are mosting likely to have greater than one viewpoint and wishes for the future. At a particular factor we will certainly attempt to identify what is the very best for every one of them.

Daniele: I simply wish to include that in order to attain the 2050 obstacle, in my viewpoint, it is essential to have simply one organization including every little thing. So as much as this factor, various organizations have actually been concentrating on various research study issues. There were organizations concentrating just on method, others concentrating just on the equipment, our organization concentrating generally on the control and vibrant handling of the gameplay. However at the end of the day, in order to take on people, there should be just one organization bringing all these solitary facets with each other. From my viewpoint, it entirely makes good sense to maintain combining organizations with each other.

Concerning the writers

	Daniele Affinita is a PhD trainee in Artificial intelligence at EPFL, concentrating on the junction of Artificial intelligence and Robotics. He has more than 4 years of experience completing in RoboCup with the SPQR group. In 2024, he operated at Sony on domain name adjustment strategies. He holds a Bachelor’s level in Computer system Design and a Master’s level in Expert system and Robotics from Sapienza College of Rome.
	Vincenzo Suriani gained his Ph.D. in Computer System Design in 2024 from Sapienza College of Rome, with an expertise in expert system, robot vision, and multi-agent control. Given that 2016, he has actually functioned as Software program Growth Leader of the Sapienza Football Robotic Group, adding to significant robot competitors and worldwide campaigns such as EUROBENCH, SciRoc, and Tech4YOU. He is presently a Study Other at the College of Basilicata, where he concentrates on establishing smart settings for software program screening automation. His research study, acknowledged with prize-winning documents at the RoboCup International Seminar (2021, 2023, 2025), fixates robot semantic mapping, things acknowledgment, and human– robotic communication.
	Domenico Daniele Bloisi is an associate teacher of Expert system at the International College of Rome UNINT. Formerly, he was associate teacher at the College of Basilicata, assistant teacher at the College of Verona, and assistant teacher at Sapienza College of Rome. He got his PhD, master’s and bachelor’s levels in Computer system Design from Sapienza College of Rome in 2010, 2006 and 2004, specifically. He is the writer of greater than 80 peer-reviewed documents released in worldwide journals and seminars in the area of expert system and robotics, with a concentrate on photo evaluation, multi-robot control, aesthetic understanding and info combination. Dr. Bloisi carries out research study in the area of cancer malignancy and dental cancer avoidance with automated clinical photo evaluation in cooperation with customized clinical groups in Italy. On top of that, Dr. Bloisi is WP3 leader of the EU H2020 SOLARIS job, device leader for the PRIN PNRR RETINA job, device leader for the PRIN 2022 AIDA job. Given that 2015, he is the group supervisor of the SPQR robotic football group joining the RoboCup globe competitors
	Can Lin is a master trainee in Information Scientific research at Sapienza college of Rome. He holds a bachelor level in Computer technology and Expert system from the exact same college. He signed up with the SPQR group in September of 2024, concentrating on jobs connected to computer system vision.