Self-supervised studying for soccer ball detection and past: interview with winners of the RoboCup 2025 finest paper award

Presentation of the perfect paper award on the RoboCup 2025 symposium.

An essential side of autonomous soccer-playing robots issues correct detection of the ball. That is the main target of labor by Can Lin, Daniele Affinita, Marco Zimmatore, Daniele Nardi, Domenico Bloisi, and Vincenzo Suriani, which received the perfect paper award on the current RoboCup symposium. The symposium takes place alongside the annual RoboCup competitors, which this 12 months was held in Salvador, Brazil. We caught up with a few of the authors to seek out out extra in regards to the work, how their methodology might be transferred to functions past RoboCup, and their future plans for the competitors.

May you begin by giving us a quick description of the issue that you just have been attempting to unravel in your paper “Self-supervised Characteristic Extraction for Enhanced Ball Detection on Soccer Robots”?

Daniele Affinita: The primary problem we confronted was that deep studying typically requires a considerable amount of labeled information. This isn’t a significant downside for widespread duties which have already been studied, as a result of you may normally discover labeled datasets on-line. However when the duty is extremely particular, like in RoboCup, you want to gather and label the info your self. Which means gathering the info and manually annotating it earlier than you may even begin making use of deep studying. This course of just isn’t scalable and calls for a big human effort.

The thought behind our paper was to cut back this human effort. We approached the issue by self-supervised studying, which goals to be taught helpful representations of the info. In spite of everything, deep studying is actually about studying latent representations from the obtainable information.

May you inform us a bit extra about your self-supervised studying framework and the way you went about creating it?

Daniele: Initially, let me introduce what self-supervised studying is. It’s a means of studying the construction of the info with out accessing labels. That is normally completed by what we name pretext duties. These are duties that don’t require specific labels, however as an alternative exploit the construction of the info. For instance, in our case we labored with photographs. You may randomly masks some patches and prepare the mannequin to foretell the lacking components. By doing so, the mannequin is compelled to be taught significant options from the info.

In our paper, we enriched the info by utilizing not solely uncooked photographs but in addition exterior steerage. This got here from a bigger mannequin which we confer with because the instructor. This mannequin was skilled on a distinct job which is extra normal than the goal job we aimed for. This manner the bigger mannequin can present steerage (an exterior sign) that helps the self-supervision to focus extra on the precise job we care about.

In our case, we needed to foretell a good circle across the ball. To information this, we used an exterior pretrained mannequin (YOLO) for object detection, which as an alternative predicts a unfastened bounding field across the ball. We are able to arguably say that the bounding field, a rectangle, is extra normal than a circle. So on this sense, we have been attempting to make use of exterior steerage that doesn’t remedy precisely the underlying job.

Overview of the info preparation pipeline.

Have been you in a position to check this mannequin out at RoboCup 2025?

Daniele: Sure, we deployed it at RoboCup 2025 and confirmed nice enhancements over our earlier benchmark, which was the mannequin we utilized in 2024. Particularly, we observed that the ultimate coaching requires a lot much less information. The mannequin was additionally extra sturdy beneath completely different lighting situations. The difficulty we had with earlier fashions was that they have been tailor-made for particular conditions. However after all, all of the venues are completely different, the lighting and the brightness are completely different, there is perhaps shadows on the sphere. So it’s actually essential to have a dependable mannequin and we actually observed an important enchancment this 12 months.

What’s your crew title, and will you discuss a bit in regards to the competitors and the way it went?

Daniele: So our crew is SPQR. We’re from Rome, and now we have been competing in RoboCup for a very long time.

Domenico Blois: We began in 1998, so we’re one of many oldest groups in RoboCup.

Daniele: Yeah, I wasn’t even born then! Our crew began with the four-legged robots. After which the league shifted extra in direction of biped robots as a result of they’re more difficult, they require steadiness and, general it’s tougher to stroll on simply two legs.

Our crew has grown loads throughout current years. We’ve been following a really optimistic pattern, going from ninth place in 2019 to 3rd place on the German Open in 2025, and we obtained 4th place at RoboCup 2025. Our current success has attracted extra college students to the crew. So it’s type of a loop – you win extra, you entice extra college students, and you’ll work extra on the challenges proposed by RoboCup.

SPQR crew.

Domenico: I wish to add that additionally, from a analysis perspective, now we have received three finest paper awards within the final 5 years, and now we have been proposing some new developments in direction of, for instance, using LLMs for coding (as a robotic’s behaviour generator beneath the supervision of a human coach). So we are attempting to maintain the open analysis discipline lively in our crew. We wish to win the matches however we additionally wish to remedy the analysis issues which can be sure along with the competitors.

One of many essential contributions of our paper is in direction of using our algorithms outdoors RoboCup. For instance, we are attempting to use the ball detector in precision farming. We wish to use the identical strategy to detect rounded fruits. That is one thing that’s actually essential for us; to exit the context of Robocup and to make use of Robocup instruments for brand new approaches in different fields. So if we lose a match, it’s not an enormous deal for us. We wish our college students, our crew members, to be open minded in direction of using RoboCup as a place to begin for understanding teamwork and for understanding how you can take care of strict deadlines. That is one thing that RoboCup may give us. We attempt to have a crew that’s prepared for each sort of problem, not solely inside RoboCup, but in addition different kinds of AI functions. Profitable just isn’t all the things for us. We’d desire to make use of our personal code and never win, than win utilizing code developed by others. This isn’t optimum for reaching first place, however we wish to educate our college students to be ready for the analysis that’s outdoors of RoboCup.

You stated that you just’ve beforehand received two different finest paper awards. What did these papers cowl?

Domenico: So the final two finest papers have been type of visionary papers. In a single paper, we needed to present an perception in how you can use the spectators to assist the robots rating. For instance, for those who cheer louder, the robots are likely to kick the ball. So that is one thing that’s not truly used within the competitors now, however is one thing extra in direction of the 2050 problem. So we wish to think about how it is going to be 10 years from now.

The opposite paper was referred to as “play in all places”, so you may, for instance, play with several types of ball, you may play outdoors, you may even play and not using a particular aim, you may play utilizing Coca-Cola cans as goalposts. So the robotic has to have a normal strategy that’s not associated to the precise discipline utilized in RoboCup. That is in distinction to different groups which can be very particular. We’ve a distinct strategy and that is one thing that makes it tougher for us to win the competitors. Nevertheless, we don’t wish to win the competitors, we wish to obtain this aim of getting, in 2050, this match between the RoboCup winners and the FIFA World Cup winners.

I’m all in favour of what you stated about transferring the tactic for ball detection to farming and different functions. May you say extra about that analysis?

Vincenzo Suriani: Our lab has been concerned in some completely different initiatives referring to farming functions. The Flourish challenge ran from 2015 – 2018. Extra lately, the CANOPIES challenge has focussed on precision agriculture for everlasting crops the place farmworkers can effectively work along with groups of robots to carry out agronomic interventions, like harvesting or pruning.

We’ve one other challenge that’s about detecting and harvesting grapes. There’s a large effort in bringing information again from RoboCup to different initiatives, and vice versa.

Domenico: Our imaginative and prescient now could be to concentrate on the brand new era of humanoid robots. We participated in a brand new occasion, the World Humanoid Robotic Video games, held in Beijing in August 2025, as a result of we wish to use the platform of RoboCup for different kinds of functions. The thought is to have a single platform with software program that’s derived from RoboCup code that can be utilized for different functions. When you have a humanoid robotic that should transfer, you may reuse the identical code from RoboCup as a result of you need to use the identical stabilization, the identical imaginative and prescient core, the identical framework (kind of), and you’ll simply change some modules and you’ll have a totally completely different sort of utility with the identical robotic with kind of the identical code. We wish to go in direction of this concept of reusing code and having RoboCup as a check mattress. It’s a very powerful check mattress, however you need to use the leads to different fields and in different functions.

Trying particularly at RoboCup, what are your future plans for the crew? There are some large adjustments deliberate for the RoboCup Leagues, so may you additionally say how this would possibly have an effect on your plans?

Domenico: We’ve a really sturdy crew and a few of the crew members will do a PhD within the coming years. One in all our targets was to maintain the scholars contained in the college and the analysis ward, and we have been profitable on this, as a result of now they’re very passionate in regards to the RoboCup competitors and about AI generally.

By way of the adjustments, there might be a brand new league inside RoboCup that could be a merger of the usual platform league (SPL) and the humanoid kid-size league. The humanoid adult-size league will stay, so we have to determine whether or not to hitch the brand new merged league, or transfer to adult-sized robots. In the meanwhile we don’t have too many particulars, however what we all know is that we’ll go in direction of a brand new period of robots. We acquired robots from Booster and we at the moment are buying one other G1 robotic from Unitree. So we are attempting to have a whole household of recent robots. After which I feel we’ll go in direction of the league that’s chosen by the opposite groups within the SPL league. However for now we are attempting to prepare an occasion in October in Rome with two different groups to trade concepts and to grasp the place we wish to go. There may also be a workshop to debate the analysis facet.

Vincenzo: We’re additionally in dialogue about the perfect dimension of robotic for the competitors. We’re going to have two completely different positions, as a result of robots have gotten cheaper and there are groups which can be pushing to maneuver extra shortly to an even bigger platform. Then again, there are groups that wish to persist with a smaller platform as a way to do analysis on multi brokers. We’ve seen loads of functions for a single robotic however not many functions with a set of robots which can be cooperating. And this has been traditionally one of many core components of analysis we did in RoboCup, and in addition outdoors of RoboCup.

There are many factors of view on which robotic dimension to make use of, as a result of there are a number of elements, and we don’t know the way quick the world will change in two or three years. We are attempting to form the principles and the situations to play for subsequent 12 months, however, due to how shortly issues are altering, we don’t know what the perfect resolution might be. And likewise the analysis we’re going to do might be affected by the choice we make on this.

There might be some adjustments to different leagues within the close to future too; the small and center sizes will shut in two years in all probability, and the simulation league additionally. So much will occur within the subsequent 5 years, in all probability greater than over the last 10-15 years. It is a essential 12 months as a result of the choices are primarily based on what we will see, what we will spot sooner or later, however we don’t have all the knowledge we’d like, so it is going to be difficult.

For instance, the SPL has an enormous, in all probability the largest, neighborhood among the many RoboCup leagues. We’ve loads of groups which can be grouping by curiosity and so there are groups which can be sticking to engaged on this particular downside with a selected platform and groups which can be attempting to maneuver to a different platform and one other downside. So even inside the identical neighborhood we’re going to have a couple of perspective and hopes for the longer term. At a sure level we’ll strive to determine what’s the finest for all of them.

Daniele: I simply wish to add that as a way to obtain the 2050 problem, in my view, it’s essential to have only one league encompassing all the things. So up up to now, completely different leagues have been specializing in completely different analysis issues. There have been leagues focusing solely on technique, others focusing solely on the {hardware}, our league focusing primarily on the coordination and dynamic dealing with of the gameplay. However on the finish of the day, as a way to compete with people, there should be just one league bringing all these single elements collectively. From my perspective, it completely is sensible to maintain merging leagues collectively.

In regards to the authors

	Daniele Affinita is a PhD scholar in Machine Studying at EPFL, specializing within the intersection of Machine Studying and Robotics. He has over 4 years of expertise competing in RoboCup with the SPQR crew. In 2024, he labored at Sony on area adaptation strategies. He holds a Bachelor’s diploma in Pc Engineering and a Grasp’s diploma in Synthetic Intelligence and Robotics from Sapienza College of Rome.
	Vincenzo Suriani earned his Ph.D. in Pc Engineering in 2024 from Sapienza College of Rome, with a specialization in synthetic intelligence, robotic imaginative and prescient, and multi-agent coordination. Since 2016, he has served as Software program Improvement Chief of the Sapienza Soccer Robotic Staff, contributing to main robotic competitions and worldwide initiatives equivalent to EUROBENCH, SciRoc, and Tech4YOU. He’s at the moment a Analysis Fellow on the College of Basilicata, the place he focuses on creating clever environments for software program testing automation. His analysis, acknowledged with award-winning papers on the RoboCup Worldwide Symposium (2021, 2023, 2025), facilities on robotic semantic mapping, object recognition, and human–robotic interplay.
	Domenico Daniele Bloisi is an affiliate professor of Synthetic Intelligence on the Worldwide College of Rome UNINT. Beforehand, he was affiliate professor on the College of Basilicata, assistant professor on the College of Verona, and assistant professor at Sapienza College of Rome. He obtained his PhD, grasp’s and bachelor’s levels in Pc Engineering from Sapienza College of Rome in 2010, 2006 and 2004, respectively. He’s the writer of greater than 80 peer-reviewed papers printed in worldwide journals and conferences within the discipline of synthetic intelligence and robotics, with a concentrate on picture evaluation, multi-robot coordination, visible notion and data fusion. Dr. Bloisi conducts analysis within the discipline of melanoma and oral carcinoma prevention by automated medical picture evaluation in collaboration with specialised medical groups in Italy. As well as, Dr. Bloisi is WP3 chief of the EU H2020 SOLARIS challenge, unit chief for the PRIN PNRR RETINA challenge, unit chief for the PRIN 2022 AIDA challenge. Since 2015, he’s the crew supervisor of the SPQR robotic soccer crew taking part within the RoboCup world competitions
	Can Lin is a grasp scholar in Information Science at Sapienza college of Rome. He holds a bachelor diploma in Pc science and Synthetic intelligence from the identical college. He joined the SPQR crew in September of 2024, specializing in duties associated to laptop imaginative and prescient.