Skild AI trains a humanoid to play soccer through 140 simulated years of self-play on top of its S1 model
Skild AI said on Sept 23 that it post-trained its S1 robot foundation model by self-play: a policy given only the goal of scoring learned to walk, get up, dribble, shield the ball and tackle by playing recent versions of itself in NVIDIA Isaac Sim for more than 140 simulated years, then transferred to a real humanoid.
The company argues that models pre-trained on human data are capped at human capability and that self-play can exceed it; it published no success rates, robot model or compute figures.
Skild AI said in a research post on Sept 23 that it had post-trained a robot policy through self-play, the method behind DeepMind's AlphaGo Zero, and transferred it from simulation to a physical humanoid that then played soccer. The policy was given one objective, to score, and learned by competing against recent versions of itself inside NVIDIA's Isaac Sim, so every gain produced a harder opponent.
The claim matters because of what Skild says it fixes. Its S1 model, launched in August, is pre-trained on human videos, glove data, simulation and teleoperation, and learns tasks from in-context demonstrations. Skild now says such models are capped at human capability, and that physical self-play is how robots will exceed it. The post shows the result on video but discloses no win rates, success metrics, humanoid model or compute budget.
What the robot learned, in Skild's description
In the first simulated months the robot could barely walk, Skild wrote. Later it learned to stand up after falling, and in time to dribble past a defender, shield the ball and tackle. None of these behaviours was rewarded directly; they emerged because they helped the policy score. After more than 140 years of simulated play, Skild transferred the policy to a humanoid and played a match against it.
Why self-play is hard to bring to robots
Self-play works when the environment is cheap to run and the score is clear, as in Go, StarCraft II and Dota 2. Robots add contact physics and the gap between simulation and reality. Soccer is a useful test because it combines balance, ball control and strategy against an adaptive opponent. Skild says it is extending the approach to everyday tasks and that its next release will look at social behaviour in larger teams, from collaborative manipulation to city-scale navigation.
The company behind it
In MEKAWIRE's verified data, Skild raised a $300m Series A at a $1.5bn valuation in July 2024 and a $1.4bn Series C led by SoftBank at $14bn in January 2026, and said on Sept 10 that it had passed $100m in annual recurring revenue with more than 60 paying customers.
- Skild AI
- Series C
- Jan 14, 2026
- $1.4bn
- $14.0bn
- Skild AI
- Series A
- Jul 9, 2024
- $300m
- $1.5bn
- Physical Intelligence
- Series B
- Nov 20, 2025
- $600m
- $5.6bn
- Figure AI
- Series C
- Sep 16, 2025
- $1.0bn
- $39.0bn
- Field AI
- Two rounds combined
- Aug 20, 2025
- $405m
- Not disclosed
Skild step-up = $14.0bn / $1.5bn = 9.3x in 18 months. Valuation-to-ARR = $14bn (Jan 2026 post-money) / $100m (ARR stated Sept 2026) = 140x. Disclosed figures only.
As of Oct 16, 2026
MEKAWIRE analysis
Skild is positioning simulation compute, not human data collection, as the scaling path for robot learning, which is a bet against the teleoperation-heavy strategies of most rivals.
The evidence is the framing: S1 for breadth from human data, self-play for capability beyond it. If it holds, the moat shifts toward whoever runs the largest physically faithful simulations, which favours NVIDIA's ecosystem and well-funded labs. The counter-argument is that soccer has a clean reward and household or factory work does not; without published metrics, one match on video does not show the method generalises. Bull case: Skild shows self-play gains on an industrial task with numbers within a year. Bear case: the approach stays a sports demo while teleoperation-data players keep winning deployments.
What to watch
Skild's promised next release on multi-agent social behaviour, any paper or benchmark numbers for the soccer policy, and whether customers in its ARR base report self-play-trained skills in production.
- More than 140 years of simulated play
- Post-training of S1 by self-play against recent versions of the policy in NVIDIA Isaac Sim
- Sim-to-real to a physical humanoid (model not named)
- Single reward: score a goal
- Skild AI research post 'Physical Self-Play', Sept 23, 2026
- Series C $1.4bn at $14bn (Jan 2026); ARR above $100m (Sept 10, 2026)
- Win or success rates, compute budget, robot model
- Standing up after falls, dribbling, shielding, tackling
Why it matters
Most robot foundation model companies scale by collecting human demonstrations, which is slow and expensive. Skild is arguing publicly that the next gains come from simulation compute through self-play. If that is right, the cost structure of robot learning shifts from labour (teleoperators) to GPUs, and the advantage moves to firms with simulation infrastructure and capital, both of which Skild has after its $1.4bn Series C.
Rival analysis
Physical Intelligence ($5.6bn post-money) and Figure ($39bn) lean on real-world and teleoperated data; Field AI targets autonomy in unstructured sites. Skild differentiates on an omni-bodied model plus a self-improvement stage. The risk for Skild is that rivals adopt self-play as a post-training step on their own data-rich bases, which would erase the differentiation.
Valuation context
At a $14bn post-money against $100m ARR stated eight months later, Skild trades at about 140x revenue on our simple method, far above industrial automation incumbents. Its 9.3x step-up from Series A in 18 months is one of the steepest in our robot foundation model data.
Supply-chain implications
Self-play at scale depends on GPU capacity and NVIDIA Isaac Sim; the robot hardware in the demo was not named. Skild's model is designed to run on many robot bodies, so it depends on partner hardware makers for deployment.
Signals to watch
1) Skild's next release on multi-agent behaviour. 2) Published metrics or a paper for the soccer policy. 3) A customer case where a self-play-trained skill is deployed in production.
Analyst view
Thesis: Skild is betting that simulation compute, not human data, scales robot capability. Confidence: low to medium, because the demonstration lacks quantitative results.
Questions you should be asking
What win rate and fall rate did the transferred policy achieve on hardware? How many GPU-hours did 140 simulated years require? Which humanoid was used? Which industrial tasks are next for self-play?