Skild AI Robot Learns Soccer From 140 Years of Simulated Self-Play

The name Messinator joins Lionel Messi's first name to the Terminator, the machine of the science-fiction film series. Its controller builds on S1, Skild AI's flagship robot foundation model, whose central ability is in-context learning: shown a video of a task, the robot treats the footage as a prompt, works out the intent behind it and converts the motions into commands suited to its own body. New tasks can therefore be attempted without retraining.

Skild AI argues that a foundation model trained mainly on human data is limited by human experience, so the company added reinforcement learning and self-play on top of S1. The robot was placed in a virtual soccer pitch built in NVIDIA Isaac Sim and given a single objective: score goals. No separate rewards were set for dribbling, shielding the ball, tackling or getting back on its feet. Its opponents were earlier saved versions of its own model, and each winning policy became the harder opponent of the next round, so every improvement generated the next problem to solve.

According to Skild AI's technical blog, the robot could barely walk in its first months on the virtual pitch. By what the company describes as college age, it had taught itself to stand up after falling. As opponents grew stronger, more complex behaviour appeared: dribbling around defenders, using its body to shield the ball and tackling to win possession. No engineer prescribed these actions; they survived only because they helped the robot win.

The approach echoes AlphaGo, which discovered moves top players had not anticipated by playing itself and beat the South Korean grandmaster Lee Sedol 4-1 in 2016. In the football demonstration, the skills were migrated to a real humanoid robot. Messinator can walk to the ball and shoot, and it also dribbles, shields, tackles and adjusts its position and movement against human opponents.

The work follows a broader line Skild AI has been pursuing. In July 2025 the company introduced Skild Brain, a robot foundation model designed to run across hardware, under the slogan "Any robot, any task, one brain." The aim is to break the industry pattern of one body paired with one bespoke brain.

Roughly two months later, Skild AI published omni-bodied training results. The team generated about 100,000 robot bodies with different structures in simulation and trained a single model for a cumulative total of about 1,000 simulated years. At test time the model had to take over robots it had never seen, including machines with a suddenly broken leg, a locked joint, a jammed wheel or extra load on its back. Instead of failing, it adjusted its centre of gravity and gait, in some cases recovering balance within seconds and working out a new way to move. The company's reasoning is that a model facing so many bodies cannot rely on memorising one fixed gait, and must instead grasp more general physics around balance, joints, centre of gravity and momentum.

Skild AI was founded in 2023 and emerged from Carnegie Mellon University's robotics research system. Its staff includes researchers from Carnegie Mellon, Stanford University and the University of California, Berkeley, along with engineers and researchers who previously worked at Meta, Tesla, NVIDIA, Google, Amazon and Everyday Robotics.

Co-founder and CEO Deepak Pathak studied computer science at the Indian Institute of Technology Kanpur and earned a computer science doctorate at UC Berkeley under Alyosha Efros and Trevor Darrell. He later researched at Meta AI Research and worked as a visiting postdoctoral researcher with the robot learning specialist Pieter Abbeel before joining Carnegie Mellon as a professor in the Robotics Institute and the Machine Learning Department. One of his best-known lines of work is curiosity-driven exploration, which gives a machine an internal drive to investigate unfamiliar and unpredictable situations rather than depending only on rewards designed by humans. His research extends to self-supervised robot learning, visual imitation learning, sim-to-real transfer and adaptive control. Before founding Skild AI, he co-founded the face recognition company VisageMap, later acquired by FaceFirst.

Co-founder and president Abhinav Gupta holds a computer science doctorate from the University of Maryland and is a professor at Carnegie Mellon's Robotics Institute. He helped establish Facebook AI Research's Pittsburgh lab and held research management roles there. His work covers large-scale visual learning, self-supervised learning, lifelong learning, robot manipulation and physical commonsense, and he was an early researcher on having robots collect training data through autonomous interaction. He has received a Sloan Research Fellowship, the PAMI Young Researcher Award and the IAPR J.K. Aggarwal Prize.