The big breakthrough in embodied intelligence is coming!!!
Since Generalist released Embodied Brain Gen 1.5 last week, capable of learning actions lasting 3 to 12 seconds～
Just now, North American embodied AI startup Skild AI unveiled its new robotic foundation model, S1, extending the context length for robotic few-shot learning to over 10 minutes.
This S1 focuses on in-context learning (ICL):
No need to specifically learn new operational data during post-training—just by watching a short human demonstration video, it can imitate and directly complete an entire complex operational workflow it has never seen before.
In the official demonstration, the robot completed long-range tasks such as making pancakes, brewing coffee, repotting plants, and assembling equipment. On unseen tasks, it achieved a 66% success rate, far surpassing language-guided VLA models (which only reached 9%).
Additionally, even if following the traditional post-training fine-tuning approach, it would take approximately 380 task demonstrations to match the performance of S1 after just one demonstration.
Very amazing!
In fact, this recent wave of work focused on in-context learning has reignited hope in embodied intelligence for many.
Some Twitter users have suggested that general-purpose AI robots could emerge in two years, not seven.
Other netizens have also noted that from Rhoda, to last week’s Generalist, and now the 10-minute tasks of Skild S1, the robot’s true GPT moment seems to be emerging.
Because once this capability is truly generalized, developing robot skills may become as simple as writing prompts for ChatGPT.
Is it really that amazing? Let’s take a look together.
From the BERT era to the GPT era
To understand how S1 is learning in this context, let’s first look at how traditional robots learn a new skill.
In traditional pipelines, robot learning is quite similar to large models:
First, pre-train on vast amounts of data to develop foundational capabilities; then, in specific scenarios, collect data tailored to a particular task and fine-tune the model to acquire specialized skills.
The issue is that although bots and large models follow similar processes, their data costs are entirely different.
The pre-training data for large models is largely sourced from the internet; even in specialized scenarios such as programming and agents, much of the post-training data can be quickly obtained online or even auto-generated.
But robots are in a tougher spot—both pre-training and post-training data still need to be collected in the real world.
So, in a technical blog, Skild AI spoke directly:
If a robotic base model still requires dozens or hundreds of hours of real-world data to learn each new task and must be retrained from scratch, then I ask: where is the “base” in this base model?
They even cite existing research indicating that, if sufficient data is available for a new task, a model trained from scratch may even match the performance of a pre-trained model followed by fine-tuning.
So, what is the significance of embodied pre-training?
In response, Skild's answer is: in-context learning.
For reference, Skild drew upon the development roadmap of large models as a benchmark.
Early BERT models were already powerful, but for each new task, you often still had to prepare new data and fine-tune again.
What truly changed the game was the emerging capability of in-context learning that became apparent after GPT-3:
You don't need to modify the model weights—just provide a few examples in the prompt, and the model can temporarily "learn" a new task.
Based on this, Skild believes that robots are currently still trapped in a BERT-like era, and what they truly want is to enable robots to undergo the same paradigm shift.
In other words, the true value of pre-training should not merely be reducing the amount of data needed for subsequent training, but enabling the model to ultimately gain the ability to learn directly from context.
Contextual learning
So, what exactly did S1 learn from the context this time?
Skild believes that there are truly only two dimensions that can test a robot's in-context learning ability:
One is whether it can learn new skills that were never seen during training; the other is whether it can combine existing skills to accomplish a long, complex new task.
For example, a robot only needs to watch a video of flipping pancakes to learn how to make them—
Even if this skill has never appeared in its training data before.
Meanwhile, compared to the second-level videos demonstrated by Gen-1.5 last week, S1 directly extends context learning to up to 10 minutes.
In tasks such as repotting plants, each task requires completing dozens of consecutive operational steps. In demonstrations of these long-range tasks, the robot must do more than simply imitate—it must understand:
Where am I now, what should I do next, how do I combine skills, and how do I recover if I make a mistake along the way?
In other words, the robot must truly understand the intent behind the tasks and actions in the video demonstration, adapting dynamically to changing situations—not merely imitating behaviors.
Additionally, in the plant repotting task, it took only 11 minutes from starting the demonstration to the robot performing the task independently, directly surpassing traditional post-training methods in efficiency.
Finally, throughout the entire process, S1 did not use fine-tuning or post-training; the model weights remained completely unchanged, and the same set of weights accomplished all the tasks demonstrated in the blog.
It has essentially bridged the gap from BERT requiring fine-tuning for specific scenarios to GPT-3 being able to perform OOD tasks based solely on demonstrations.
The only difference is that the prompt is no longer language, but rather action videos better aligned with downstream tasks.
Experiment: Is ICL truly more scalable?
In the experimental section, Skild compared the ICL video prompt with traditional language prompts for VLA on both seen and unseen tasks (from the training data).
Test results show that when the training data is only 1,000 hours, language prompts perform better, but as the data scale increases, ICL begins to outperform them.
At 100,000 hours, performance on tasks seen during training: ICL 96%, Language Prompt 89%.
On the truly critical OOD tasks: ICL achieves 66%, while language prompts only reach 9%, creating a gap of more than sevenfold.
To verify the generalizability of S1, Skild conducted additional tests under varying experimental conditions.
The results show that the performance degradation of the language prompt VLA can be up to three times that of ICL.
In other words, ICL does not simply repeat actions in fixed scenarios, but instead adapts and replans its actions based on the current environment.
Finally, in terms of one-shot learning.
S1 achieves a 66% success rate on a new task with no post-training, relying solely on a single video demonstration.
In comparison, traditional VLA requires approximately 380 post-training demonstrations to reach the same level.
Of course, after continuing to stack up to 2,000 demonstrations, traditional post-training eventually achieves 86%.
So the conclusion may not be "robots won't need training anymore," but rather:
Previously, a new skill might require hundreds of lessons to learn, but now, after just one look, you can immediately achieve a score of 60 to 70.
This is also what Skild refers to as the ICL scaling law:
The more data, the more the model doesn't just acquire more skills, but learns to acquire skills from demonstrations.
Who is Skild AI?
Skild was founded in 2023 by two familiar faces from the CMU robotics community: Deepak Pathak and Abhinav Gupta.
Both are professors at the Robotics Institute of Carnegie Mellon University. Deepak primarily researches robot learning, reinforcement learning, and computer vision, while Abhinav is a leading expert in computer vision and self-supervised learning.
As early as 2022, the two collaborated on WHIRL, enabling robots to perform one-shot imitation by watching human videos—so this S1 approach didn’t just emerge out of nowhere.
As a standout company in North America's embodied AI circle, Skild emerged from stealth mode in 2024 and secured a $300 million Series A round at a $1.5 billion valuation;
By January of this year, the company completed a $1.4 billion Series C round, pushing its valuation above $14 billion, led by SoftBank, with NVIDIA, Bezos, and others continuing to participate. In just over two years, its valuation nearly doubled tenfold.
From a horizontal perspective, Skild holds a uniquely distinctive position in the North American embodied AI community.
Compared to PI, which seems more focused on "robotic GPT," Generalist has recently emphasized being a one-shot learner, while Genesis AI and Sunday have shown strong full-stack momentum. Skild has consistently emphasized one phrase: Any robot, any task, one brain.
It places greater emphasis on cross-embodiment, aiming for the same Skild Brain to run on different bodies such as robotic arms, quadrupeds, and humanoids.
In simple terms, it’s like wanting to be the Android of the robot era—a smart layer that can be plugged into all kinds of different bodies.
This also determines Skild’s data strategy: everything.
Skild views robot data through three dimensions: hardware proximity, diversity, and scalability:
Real hardware remote control is closest to actual hardware but is expensive; first-person human video is easiest to scale but differs greatly from robotic bodies; simulation is cheap and can be rapidly produced, yet suffers from the sim-to-real gap.
Therefore, they essentially adopt a strategy of using all of them for Robot Teleop, UMI, Egocentric Video, and Simulation.
And S1's ICL is also a natural extension of this approach:
September 2025: LocoFormer → February 2026: First In-Domain ICL → May: First flip of the pancake → August: S1 release, directly pushing in-context learning to OOD long-range tasks of up to 10 minutes.
So the real question worth focusing on now may no longer be whether robots can learn more skills, but whether:
Can the cost of teaching a robot a new skill really shift from "teaching it hundreds of times" to "you do it once, and it watches once"?
If this scaling law continues to hold, the so-called GPT moment for robots might truly be more than just another marketing buzzword.
Reference link: [1] https://www.skild.ai/blogs/s1
This article is from the WeChat public account "Quantum Bit," authored by Henry.