what is egogoal-amd-hackathon fr?

MESSI, short for Motion Estimation for Soccer Skill Imitation, is a research project that teaches a humanoid robot to kick a soccer ball the way a human athlete does, by learning from video of real people kicking. It takes 3D joint positions extracted from human video, maps them onto the joint structure of the Unitree G1 humanoid robot, and then fine tunes the resulting movement so the robot's kick actually works under realistic physics, including how a ball rolls and slows down on turf. The project explains that most humanoid robots are good at steady tasks like walking or standing, but struggle with fast, high force athletic moves such as kicking or jumping, because those moves need precise timing, high joint speeds, and careful balance during the moment of contact. Simply copying a human's joint angles onto a robot does not work well, since people and robots have very different body proportions and joint limits, so an untuned kick barely moves the ball, in this case only about 3.7 meters, far short of the roughly 13 meters needed to reach a goal. To fix this, the system uses a two stage approach. First it maps the human's joint movements onto the robot using the robot's own hip position as a reference point, preserving the shape of the original kicking motion while adapting it to the robot's body. Then it runs an optimization process that adjusts the leg and body motion in detail, using fast gradient based tuning combined with a search algorithm that tries many candidate movements in a physics simulator and keeps the best ones. After this optimization, the robot's kick reaches ball speeds up to 9.9 meters per second and travels over 13 meters, enough to score. The heavy computation involved in searching through many candidate movements is sped up using AMD's ROCm software stack running on AMD Instinct MI300X graphics processors, which lets the system evaluate dozens of candidate kicks per generation and find a stable, powerful kicking motion in under two minutes.