The race for generalist robots has a new benchmark to beat. Skild AI launched S1 on August 25 and 26, 2026.
Industry observers are calling it the GPT-3 moment for physical AI. S1 is a foundation model built for robot control.
It learns new tasks from a single video of a human demonstration. The robot then executes that task for up to ten minutes continuously.
It handles actions it was never explicitly trained to do. That kind of in-context learning has never worked this well for robots.
Alfred Lin of Sequoia called the single-prompt ten-minute execution a genuine game changer. Skild is valued at over fourteen billion dollars following the announcement.
What the Benchmarks Show
According to Skild’s own announcement, S1 achieves a 66% success rate on unseen long-horizon tasks. Rival systems prompted only with language score just 9% on the same benchmark.
That is a seven-fold improvement in one generation of physical AI. Those numbers demand careful and honest scrutiny from the field.
The benchmark comes entirely from Skild’s internal evaluation team. No independent third-party replication has been published yet.
Model weights have not been released to outside researchers either. The gap is striking, but the full scientific picture is still forming.
Generalist AI also shipped its GEN-1.5 model one day earlier, on August 24. The physical AI foundation model race is now a real and active competition.
Both announcements arrived within 48 hours of each other.
From Lab to Warehouse Floor
S1 has real industrial partners, not just research collaborators. Skild has already signed deployment agreements with ABB and Universal Robots.
Both companies operate robots in live manufacturing and logistics environments globally. That means S1 could reach active warehouse floors within months, not years.
Skild also expanded its capabilities through strategic acquisition. In April 2026, it acquired Zebra Technologies’ robotics division, Zebra Robotics.
That deal added proven hardware expertise and an existing commercial install base. Anthropic’s Model Hardware Standard is part of the same physical AI infrastructure push.
The race to give AI agents control over physical systems is now accelerating broadly. The bottleneck is now hardware deployment speed and real-world safety validation.
What This Means for the Industry
S1 shifts the standard expectation for what a robot should do. Previously, industrial robots required weeks of programming per new task.
One video and ten minutes of autonomous execution changes that calculus entirely. The benchmark caveats still matter for buyers and investors.
Internal evaluations can reflect controlled best-case conditions only. Real warehouses bring noise, lighting variation, and unexpected obstacles.
Field performance will be the true test of S1’s claims. Even so, the direction is unmistakably clear.
Generalist robot models trained on video are closing in on practical utility. The next twelve months will test whether S1’s lab numbers hold up.
If they do, the physical AI moment is no longer hypothetical.
