Team Advanced Robotics Laboratory (AdRoLab) from Unicamp, in partnership with the Hub of Artificial Intelligence and Cognitive Architectures (H.IAAC), won first place in the PhyRC Challenge. This is an international competition in the area of assistive robotics, which uses robots to help people with some type of limitation in routine tasks.
Competing among 56 universities from 17 countries, Brazil was the only country from the Global South to participate in the competition. This is the first time that a team from Unicamp has competed, alongside internationally renowned institutions such as Stanford University, Yale University, Cambridge University, among others.
O PhyRC Challenge – organized by EmPRISE Lab (Cornell University laboratory, United States) and sponsored by the company Kinova – had two stages. The first consisted of a virtual simulation of the process of putting a lab coat on a mannequin using only a robotic manipulator. The second proposed the challenge of putting a shirt on a real mannequin during the IEEE International Conference on Robotics and Automation 2025, the largest international event in the field of robotics, held in May in the United States. “It is a big challenge, even for a human being it is difficult [to put on a shirt using only one arm]. This physical part of artificial intelligence is complicated, because the technology is not yet ready to solve this type of situation”, said Eric Rohmer, professor at the School of Electrical and Computer Engineering (Feec), researcher at H.IAAC and leader of AdRoLab.
The result came as a surprise to the team, who struggled to find a promising approach. After exploring the latest machine learning techniques (“machine learning”) unsuccessful, the solution was to bet on a hybrid proposal, through the Gaussian Mixture Model (GMM) – mixing modern robotics based on artificial intelligence (AI) and the previous robotic methodology, based on mathematical models.
The difference lies in training the robot to perform the task. While AI requires a large volume of data, with MMG, five demonstrations were enough for the robot to be able to perform the action. “We run an algorithm that reads the position of the manipulator while we literally grab the shirt and make the manipulator put it on. We feed these demonstrations into MMG, which models this data,” explained Maria Fernanda Paulino Gomes, a master’s student who had the idea of applying the technique and will base her dissertation on it.
When putting on the shirt, the model analyzes the robot's position and uses probability to determine which path has the highest chance of success, using data from all the trajectories learned. This innovative solution was what guaranteed the highest score for the Brazilian team, which, during the competition, managed to put on one arm and the head of the mannequin.
The methodology also made it possible to replicate the conditions of the tests carried out with the mannequin in Brazil in the United States, without the need for many changes. “This technology is very promising for current robotics; it is a contribution that one would expect in a postdoctoral fellowship,” Rohmer emphasized.
For Paula Dornhofer Paro Costa, a professor at FEEC and coordinator of the Cognitive Architectures research line at H.IAAC, the ability to improvise solutions to problems was a distinguishing feature of the team. The mannequin used in the tests, for example, was adapted from a model for clothing stores. Participation in the competition was supported by the Office of the Vice-Rector for Research. “Participating in the conference gave us the opportunity to follow technological trends and understand where they are heading,” Costa emphasized.
Elton Cardoso do Nascimento, Ervin Bolivar Huayhua, Esther Luna Colombini and Leonardo Rocha Olivi (Federal University of Juiz de Fora) were also part of the winning team.
Watch the video produced for the competition:
Physical artificial intelligence
So-called “physical artificial intelligence” integrates AI into robots that can interact with the world, using sensors to read their surroundings. These intelligent robots are useful in the context of an increasingly aging and care-deprived global population. “This award recognizes the expertise from Unicamp in this area of assistive robotics, which is important and needs more investment”, highlighted Costa.
According to Rohmer, the competition challenge fits in with some of the research that AdRoLab has been developing: assistive robotics focused on mobility and manipulation, using a robotic wheelchair and a lower limb exoskeleton. The satisfactory result with the Gaussian Mixture Model motivated the laboratory to explore new research with this hybrid methodology, opening up a new range of studies. “AI is the future, but we still do things with robotics techniques from 50, 60 years ago. We worked with H.IAAC to make a smooth shift from old paradigms to more recent ones within the physical AI [Physical AI]”, defended Rohmer.
By winning the competition, AdRoLab will receive another Kinova brand manipulator. “We will continue in this direction to complete the challenge and present a possibility of dressing a disabled person, an elderly person, using a manipulator,” said the researcher.
One of the obstacles to be overcome is the machine’s ability to identify the state of the shirt – whether it is crooked, wrinkled, etc. – and make the necessary adjustments to fit the mannequin. To achieve this, the team is also working with a camera that identifies the position of the body and the clothing. “We painted the shirt [which was completely white] so the robot could detect it automatically. This was not used for the competition, but we will use it in the future,” said César Bastos da Silva, a doctoral student and member of the winning team.
The expectation is to integrate the manipulator with other robots focused on mobility, placing it on top of a wheelchair, for example. “We want the wheelchair user to be able to move around, but also interact with everyday objects. And, when the person is not in the wheelchair, we can use it as a service robot, which will help them with tasks such as getting dressed or washing the dishes in an automated way, without the need for supervision or orders from a human being,” Rohmer explained.
