Imagine showing a robot how you pot a plant, then asking it to take over. No programming lesson. No painstaking list of instructions for every movement. Just a demonstration of the job you want done.
That is the direction Skild AI is pursuing with S1, a model highlighted in a 10 September 2026 Nvidia update. Nvidia says demonstrations include preparing coffee, making pancakes and assembling kits, with some tasks lasting up to ten minutes. These are company-reported demonstrations, not a guarantee of dependable performance in your kitchen. Nvidia’s S1 update
A video becomes the instruction
Skild describes the approach as in-context learning. The demonstration is supplied as an example when the robot is asked to work; the model does not need its underlying parameters rewritten for that particular task. Those parameters are the numerical settings learned during training. Skild’s technical report
Think of a capable colleague watching how you pack an order. The demonstration explains the local routine. It does not teach the colleague everything they know about objects, movement or packing from scratch.
The same distinction matters here. “One video” does not mean “one video’s worth of learning in total”. Nvidia describes a much larger development process involving simulation, human video, remotely operated robots and other training experience. Its Cosmos, Isaac and Omniverse technologies help supply and test that experience. Nvidia’s development overview
The headline number is not a finished-job score
On Skild’s internal unseen-task benchmark, the reported result was 66% versus 9% for a language-prompted comparison system at the same pretraining-data scale. Crucially, the measure was average cumulative per-step success. It was not the percentage of complete jobs finished flawlessly. The evaluation also used human interventions to recover failed attempts so later steps could be graded, mainly for the comparison system. Benchmark methods and results
That makes the result interesting without making it a household reliability rating. For an owner, a robot that completes most movements but leaves the coffee unfinished can still represent an unsuccessful job.
The next useful evidence would be independent testing: unfamiliar objects, awkward lighting, interrupted routines and repeated shifts of work. Reports should count how often a person has to step in, not just how many actions look successful.
The business opportunity is changing the job cheaply
The most consequential application may be less glamorous than a robot chef. Consider a small manufacturer whose packaging changes every few weeks. If staff could demonstrate a revised routine quickly, automation might become practical for work that changes too often to justify elaborate reprogramming.
That is a potential benefit, not an established saving from these experiments. Hardware costs, safe operation, supervision and maintenance still belong in the calculation.
It also helps explain Nvidia’s interest. The opportunity reaches beyond the robot itself to the computing and simulation used to train it. Our earlier feature on why robot hands are harder to build than robot brains explores the physical side of that challenge.
S1 offers a compelling glimpse of a more teachable machine. The milestone to watch next is whether a robot can turn “watch me do this” into a repeatable working day.
Leave a Reply