The aim of the competition is to specifically advance research in the field of robotic athletic intelligence. The participating teams develop a global control policy to solve the so-called swing-up problem on an underactuated two-joint robot system. The robot is to be transferred from any initial state to an upright position - first in the simulation and then on real hardware. This also requires a robust design that can withstand external disturbances.
The competition is aimed in particular at students and researchers from the fields of artificial intelligence, machine learning, reinforcement learning, optimal control and related disciplines.
The following teams, who have already successfully solved the swing-up problem in the simulation, were selected for the final phase:
- Team AR-EAPO - Korea University & Independent Researcher (South Korea)
Average-Reward Maximum Entropy Reinforcement Learning for Global Policy in Double Pendulum Tasks - Team MC-PILCO - University of Padova (Italy) & Mitsubishi Electric Research Laboratories (USA)
Learning Global Control of Underactuated Double Pendulum with Model-Based Reinforcement Learning - Team NMPC - University of Bremen & DFKI (Germany)
Real-Time Model Predictive Control for the Swing-Up Problem of an Underactuated Double Pendulum - Team AORRT - Rutgers University (USA)
PRACSYS Solution for 3rd AI Olympics @ ICRA 2025 - Team VIMPPI - KAIST (South Korea) & Innopolis University (Russia)
Enhancing Model Predictive Path Integral Control with Variational Integration for Underactuated Systems
The winning teams will receive prizes totaling 2,500 US dollars and have the opportunity to present their results to an international audience of experts at ICRA 2025.
Further DFKI research contributions at the ICRA 2025:
In addition to organizing the AI Olympics, the DFKI Robotics Innovation Center will be represented at the conference with the following scientific papers:
Franek Stark, Jakob Middelberg, Dennis Mronga, Shubham Vyas, Frank Kirchner: Benchmarking Different QP Formulations and Solvers for Dynamic Quadrupedal Walking, Accepted at ICRA 2025, DOI: 10.48550/arXiv.2502.01329
Maximilian Albracht, Shivesh Kumar, Shubham Vyas, Frank Kirchner: Model Predictive Parkour Control of a Monoped Hopper in Dynamically Changing Environments, Accepted at RA-L with ICRA presentation, DOI: 10.1109/LRA.2024.3445668
Further information: https://ai-olympics.dfki-bremen.de
Contact:
Dr. Dennis Mronga
Robotics Innovation Center
German Research Center for Artificial Intelligence GmbH (DFKI)
E-Mail: dennis.mronga[at]dfki.de