AlphaGo, but for humanoid soccer. Skild AI brings self-play to humanoids. Their flagship robotics foundation model, S1, can now be post-trained with no human demonstrations by competing against itself in simulation. The testbed is soccer, which demands balance, agility, ball control, reactivity, and strategy. Given a single objective (score goals), S1 played 140+ years in NVIDIA Isaac Sim and went from falling over to dribbling past defenders, shielding, tackling, shooting, and getting back up. None of these behaviors were hand-rewarded; they emerged because they helped it score. Its opponents (past versions of itself) improved alongside it, creating an automatic curriculum. With 4 agents, passing and coordination started to emerge. Skild says the approach extends to social navigation, collaborative manipulation, and any multi-agent task with a simulator and an objective.

Humanoid football had a big year. For comparison, this is what the world finals of RoboCup 2026 looked like just last month in South Korea. Should we compete next year? ;)

140 years of soccer practice in simulation. ⚽ Skild AI’s robot policy learned to dribble, shield, and tackle through self-play in NVIDIA Isaac Sim before being transferred to a real humanoid robot. 👇

This is one of the most fun projects I've worked on. Playing against the robot was surreal. It's surprisingly fast! We tried many approaches, but the simplest worked best at scale: self-play on top of a skill-conditioned low-level policy.

Skild's S1 stands out because it showed more pretraining data made the model substantially better at performing unseen tasks from a single video prompt, without fine-tuning. A stronger foundation model made each new demonstration more useful. Now, Skild is showing a complementary approach: generating experience that humans never have to demonstrate. Skild post-trained a soccer policy through 140 years of simulated self-play against earlier versions of itself. No human demonstrations during this post-training phase. Behaviors like shielding the ball and tackling emerged without separate rewards engineered for each skill. The policy then transferred to a physical humanoid. AlphaGo was such an important moment for AI because human expertise could be a starting point rather than the ceiling. A system could discover strategies nobody had explicitly taught it. Self-play creates a powerful learning loop. As the model improves, it becomes a tougher opponent for its next version. Progress generates the next challenge. More compute can produce new experience, rather than just more training on a fixed dataset. What’s interesting here is applying it on top of a general robotics foundation model, with behaviors transferring into the physical world. Together, Skild’s results suggest two complementary ways to scale: learn more effectively from the experience humans provide, then improve through experience the model generates itself. Soccer has clear objectives and measurable outcomes. Extending this to everyday work, and showing that the gains generalize beyond a single task, is the next test. Skild is showing how a capable foundation model could help change that constraint of human capability: human demonstrations get it started, without having to supply every subsequent improvement.

Very cool result, Congrats! There is an insightful quote from philosopher Merleau-Ponty, " We know not through our intellect but through our experience" . The essence of robot learning is with reinforcement learning, as it is with humans.

This robot can actually take you on at soccer now. ⚽ Skild AI is fine-tuning its S1 model with self-play reinforcement learning, the idea that helped power AlphaGo and OpenAI’s Dota 2 bots. S1 builds on skills it already has, playing recent versions of itself in simulation with one goal: score. As it improves, newer versions become tougher training opponents. After 140+ simulated years, it learned to shield the ball and tackle without separate rewards for those moves. They emerged because they helped it score. No new human demonstrations are needed during self-play. In early four-agent simulations, passing and coordination are starting to emerge, too.

⚽️ 🤝 🤖 @SkildAI is bringing back self-play, this time for physical AI. Inspired by the World Cup this summer, Skild started from S1 to have a humanoid play soccer against increasingly capable versions of itself: - for 140+ simulated years - only objective during self-play: score goals No new demonstrations. No hand-crafted rewards for dribbling, shielding, tackling, shooting, or recovering from falls. Those behaviors *emerged* because they helped it win, and transferred to a real robot A scalable way for robots to create their own curriculum, improve beyond human-provided data, and keep making themselves better!

A month ago, S1 showed it could learn a new task from a single video. Now it can post-train through self-play without new demonstrations. Another step toward robots that can keep learning beyond human-provided data. Incredible progress from @SkildAI, @deepakpathak and @gupta_abhinav_

In the race to build better AI models, most of the attention has gone to pre-training. But in robotics, we believe post-training is just as critical. Pre-training can get you impressive capabilities. Post-training is what closes the gap between a demo and a deployment—between a video and actual dollars. In fact, the real frontier in robotics may increasingly lie in post-training. Today, we’re introducing a new approach to post-train robotic policies using self-play. Inspired by the original self-play work at DeepMind, as well as our own work on robust adversarial reinforcement learning, we train policies in simulation to help emerge behaviors and make them robust before they ever reach the real world. This is an early step (and on sports), but we believe it points toward something much bigger: bringing some of the original ideas that made reinforcement learning so powerful back into the physical AI stack.

For real world robotics deployments, learning the initial procedure is only part of the challenge. Most startups' robot foundation models are already able to do that. To deploy in to the real world, robot also needs to adapt when that procedure changes. In-context learning offers a way to communicate those changes through another demonstration, rather than automatically starting a new data-collection and training project. That reduces the engineering required to keep complex deployments useful as the customer’s operation evolves. Skild had a unique focus on in-context learning, and it is reaping the rewards in the scale of their deployments within the first year.

The robotics companies that win won’t just have the best demos. They’ll have the strongest deployment loop. Skild AI: first commercial deployment → 60+ paying customers → $100M ARR in 10 months. Every deployment creates new problems to solve and new knowledge to feed back into the system.

The best place to learn isn't in the lab. It's in the real world, in actual deployments with actual customers. Skild has shown off their impressive tech before, but they're now also showing off their impressive commercial traction and have already passed 100M ARR. If we haven't reached the "ChatGPT moment for physical AI" yet, we're surely close.