Skild Humanoid Robot Learns to Play Soccer Through Self-Play
Sequoia partner Stephanie Zhan shared and introduced that Skild AI brings self-play back to physical AI: starting from the foundational model S1, humanoid robots play soccer against increasingly stronger versions of themselves in simulations, then transfer the strategies to real machines.
The company blog states that this is the post-training result for S1. Training was conducted in NVIDIA Isaac Sim with the sole goal of scoring; there were no new human demonstrations, nor were rewards separately designed for dribbling, shielding, tackling, shooting, or getting up after falling. These actions emerged on their own as they contributed to winning. In the first few months of simulation, the robot could hardly walk, but by 'college age' it could fall and get back up; after accumulating over 140 simulated years, the strategies were transferred to the real machine for competition. The company describes the transfer from simulation to real machine and the complexity of tasks as clear signals on the path to physical general intelligence.
S1 is set to be released in August 2026, positioned as a cross-modal foundational model for robots: a single video demonstration can execute unseen operations lasting up to ten minutes, with weights kept frozen. The company claims that the average success rate for unseen tasks is about 66%, with a language prompt baseline of about 9%. The data recipe mixes simulation, internet human videos, claw collection, and a small amount of remote operation. Skild was founded in 2023 in Pittsburgh by co-founders Deepak Pathak and Abhinav Gupta, with public financing totaling nearly $1.7 billion; a $300 million round in 2024, participated by Sequoia, valued the company at $1.5 billion, with Zhan joining the board.
Zhan cites the World Cup as an inspiration, emphasizing that scalable curricula are generated by the robots themselves, with capabilities potentially exceeding human demonstrations. The official inquiry is: what would happen if the same self-play were applied to virtual construction sites, factories, or homes and scaled up by several orders of magnitude? The company states it is already expanding this method to everyday tasks outside of sports. Existing robot soccer competitions are still far below human levels, with soccer chosen as a testing ground for both physical and strategic challenges.
In market mechanisms, this is a narrative of capabilities, not an immediate purchase order. Buyers are automotive manufacturers, logistics, and humanoid hardware vendors seeking transferable, minimally labeled intelligent systems; sellers face pressure to collect data by task and retrain strategies by model type in traditional robot software. Funding has been priced in rounds as 'one brain running multiple bodies.' Beneficiaries are simulation stacks and foundational model laboratories, while project-based integrators that must write reward functions for each action are under pressure.
On a supplementary level, self-play has been pioneered by DeepMind in Go and games; here, the challenges are continuous torque, balance, and contact mechanics, rather than discrete moves. The company has not yet disclosed quantitative results such as the real machine's scoring rate against humans or other robots, duration of confrontations, and frequency of falls.
Source: Public Information
ABAB AI Insight
Skild is not selling humanoid hardware, but transferable strategies. Pathak and Gupta emerged from Carnegie Mellon-style embodied research, and Sequoia priced the company as 'GPT-3 for the physical world' during its formation. S1 first proves that a single video can drive ten minutes of operation, and self-play further demonstrates that post-training can no longer rely on human demonstrations. 140 simulated years have shifted data collection from workshops to GPU calendars, which is the scale economy in the story of rounds.
Capital bets on curriculum generation after the collapse of reward functions. Scoring is a sparse goal, while dribbling and tackling become emergent skills, meaning that in a factory, 'putting parts into bins' can also be set to only define the final state, allowing strategies to develop grabbing and obstacle avoidance on their own. Whoever first moves self-play from the soccer field to the production line can transform the data moat from remote-operated farms to simulation clusters. Sequoia is simultaneously investing in chip design and open-source superintelligence; Skild writes the same set of 'scale produces intelligence' into the body.
Similar structures include AlphaZero for Go, OpenAI Five for games, and quadrupedal parkour in simulations. The difference is that bipedal ball control binds contact, vision, and confrontation into the same strategy, where transfer failures can lead to direct crashes. The industry is at a stage where operational foundational models can just learn context, with sports becoming a public demonstration arena. Generalist and other laboratories focus on short-term operations, while Skild uses soccer to write dynamic confrontation into the next report card.
Essentially, this is a technological replacement. The mechanism is: the bandwidth of human demonstrations is limited, and self-play opponents can continuously grow stronger; when the only reward is the final win or loss, intermediate actions no longer need to be named by engineers. The data flywheel shifts from 'humans teaching robots' to 'robots challenging themselves,' and hardware companies will become peripherals for models.
ABAB News · Cognitive Laws
When only rewarding goals, dribbling will emerge on its own.
As opponents grow stronger synchronously, curricula no longer need to be handwritten.
After simulating a century, the real machine only begins to resemble one that can play soccer.