Unveiled by Skild AI in August 2026, S1 is a foundation model for robotics designed to perform a new task from a single video demonstration, without any additional task-specific training. The published trials show progress on manipulation sequences lasting up to ten minutes, but they still rely on internal evaluations. The stakes therefore go beyond the technical demonstration: the question is whether this in-context learning can become reliable, safe and controllable enough for professional environments.
What Just Happened
Skild AI unveiled S1 in August 2026 as its flagship foundation model for robotic manipulation. The system receives a video showing a task, then attempts to reproduce the observed goal with a robot, without changing its weights and without any task-specific post-training phase. On September 10, NVIDIA detailed the infrastructure used to train, simulate and deploy the model, including Cosmos, Omniverse, Isaac Sim and Isaac Lab.[1][2]
The promise concerns new, long-horizon tasks. Skild AI shows S1 repotting a plant, brewing filter coffee, making crêpes or assembling a kit. Some sequences last up to ten minutes and involve several successive movements. For the repotting task, the company says it went from recording the demonstration to the start of autonomous execution in eleven minutes.[1][2]
This timeline needs clarifying: S1 was not launched on September 10. The initial announcement dates back to August 2026, while NVIDIA’s September 10 publication presents the technical collaboration and the use of its computing and simulation platforms.
What Is Actually New
The idea of teaching a robot a movement by demonstration is not new. Programming by demonstration, imitation and teleoperation have been studied for decades. What sets S1 apart is its ambition to bring to robotics a mechanism that has become central to large language models: in-context learning.
In this framework, the video is not used to retrain the model. It is placed in the model’s context at the time of use. S1 must extract the intent from it, recognize the relevant objects, track the task’s progress and translate the demonstration into actions compatible with the robot and the scene at hand. The company states that the same set of weights produces all the examples shown.[1]
The claimed novelty also concerns the length and novelty of the tasks. Skild AI says that several scenarios did not appear in the pre-training data and that they combined basic movements in new sequences. If this capability is confirmed outside internal trials, it could reduce the cost of data collection and specialization for each new operation.
How S1 Learns From a Video
S1 is trained on episodes in which a visual demonstration describes the task to be performed. The model learns to link what it observes in the video to the current scene and the available motor commands. It therefore does not necessarily copy each of the demonstrator’s movements. Instead, it seeks to reconstruct an action policy that pursues the same goal with its own robotic body.
Pre-training combines several families of data: teleoperation, egocentric human videos, simulation and experiments carried out on robots. Each addresses a different constraint. Teleoperation is close to real hardware but costly to produce. Human video brings more diversity, but with a wider gap between human and robot movements. Simulation scales more easily, but it must accurately represent contacts, pressure, collisions and physical properties.[1]
NVIDIA is involved at several stages of this pipeline. Cosmos helps organize or enrich video data, Omniverse and Isaac Sim provide virtual environments, Isaac Lab is used for reinforcement learning and TensorRT optimizes inference. These tools are not S1 itself. They form the infrastructure on which Skild AI trains and deploys its model.[2]
Announced Capabilities and Key Limitations
What the Results Actually Support
Skild AI compares a policy guided by video demonstration with a VLA (Vision Language Action) model driven by a language instruction. On new tasks lasting four to eight minutes, trained on a reported volume of 100,000 hours of data, the company reports an average cumulative per-step success rate of 66% for S1, versus 9% for the language-based baseline.[1]
This metric does not mean that 66% of each mission is fully completed. It aggregates success across the different steps. Skild AI also specifies that human interventions were used to reset the system after a failure so that all steps could be evaluated. This methodology provides information about the robot’s local progress, but it does not directly measure the proportion of tasks completed in full without assistance.
The company also estimates that one in-context demonstration achieves a level comparable to roughly 380 post-training episodes on the tasks studied. This equivalence is the result of interpolating between several experimental data points. It should not be understood as a universal rule applicable to any task, hardware or environment.[1]
These results support a limited but important conclusion: within Skild AI’s protocol, a video demonstration provides more information than a text instruction for certain long, new tasks. They do not yet demonstrate that a general-purpose robot can learn any task from a video, nor that it can operate at the level of safety required in production.
As of September 22, 2026, no peer-reviewed publication or independent reproduction of S1’s full results had been identified. The company’s videos and figures are therefore useful primary evidence, but still insufficient to conclude that it generalizes robustly.
Why This Approach Could Matter for Industry
In manufacturing and logistics, a significant share of integration costs comes from adapting robots to products, workstations and work sequences. If a video demonstration can truly reconfigure a task in a few minutes, production lines could become more flexible, particularly for small batches, frequent changeovers or operations that are difficult to formalize in code.
The role of operators would also evolve. Their hands-on expertise could become a direct source of instructions for robotic systems. This would create new needs in preparing demonstrations, validating trajectories, analyzing errors and supervising deployments. Automation would therefore not necessarily eliminate human involvement, but would shift part of the work toward transferring, controlling and governing physical skills.
Skild AI and NVIDIA cite use cases in manufacturing, logistics, inspection, security and food preparation. In particular, a partnership with Foxconn is set to apply the Skild Brain to dual-arm manipulators for assembling NVIDIA Blackwell systems.[2] These announcements are signs of deployment, but public data does not yet make it possible to assess availability, sustained reliability, cost per task or incidents encountered in production.
One Video Is Not Enough to Guarantee Generalization
A demonstration covers only one possible trajectory. In a real workshop, objects change orientation, tools wear out, lighting varies and people may enter the work area. Generalization requires the model to distinguish the goal from the exact sequence observed and to know how to respond to situations absent from the video.
Skild AI shows trials in which S1 continues a task after an object has been moved or substituted. The company also says the system can resume an action after certain failures. These behaviors are interesting, but they are still presented on selected examples. An industrial evaluation should document complete failure rates, near misses, recovery conditions, differences between robots and degradation over long periods.
Rapid adaptation also creates a validation problem. If any operator can teach a new task, the organization must decide who is authorized to do so, how the demonstration is verified, in which environment it can be executed and how to roll back to a previous version. In-context learning potentially reduces programming time, but it removes neither the need to qualify the process nor responsibility for the deployment.
Safety, Liability and Legal Framework
A robotic model acts in the physical world. An error does not just produce incorrect text: it can damage an object, halt production or injure a person. Mechanisms for force limiting, emergency stops, zone separation, speed monitoring and human validation therefore remain essential, even when the action policy is generated by a foundation model.
In the European Union, the Artificial Intelligence Act is being phased in gradually. An AI system integrated as a safety component into a product covered by EU legislation may fall under the high-risk regime when the conditions of Article 6 are met. The obligations associated with systems covered by Article 6(1) are in principle set to apply from August 2, 2027.[4] Not every robot using S1 therefore automatically becomes a high-risk system: classification depends on the function, the product and the context of use.
The EU Machinery Regulation, which is set to apply mainly from January 20, 2027, also increases the attention paid to software and to machine safety functions.[5] For a professional deployment, the assessment will need to bring together machine safety, cybersecurity, management of software changes and the requirements related to the AI system. This overview remains general and does not constitute legal advice.
Liability must also be clearly allocated among the model provider, the robot manufacturer, the integrator, the employer and the user. Execution logs, the model version, the video used as a demonstration, the safety settings and human interventions must all be traceable when a decision or incident is analyzed.
Data, Labor and Environmental Impact
Demonstration videos can capture employees, production spaces, confidential processes or protected objects. Their collection must be limited to what is necessary, secured and governed by retention and access rules. When individuals are identifiable, the GDPR may apply. Organizations must also verify their rights to the videos and the protection of their industrial know-how.
On the social front, learning by observation can showcase operators’ expertise, but it can also make it easier to capture their movements without adequate recognition. Responsible deployment requires involving teams, documenting changes to job roles, providing training in supervision and preventing video from becoming a tool for individual surveillance. Nor does the ergonomic quality of a human movement mean that it is automatically safe or optimal for a robot.
Finally, Skild AI has not published a full energy assessment for S1’s training or inference. Simulation and the reuse of a general-purpose model may reduce some physical data collection, but training on large volumes of data and relying on accelerated infrastructure carry a material and energy cost. Without comparable measurements, no net environmental benefit can be claimed.
What to Watch Next
The next decisive step will not be another demonstration video, but the publication of reproducible protocols and independent results. In particular, it will be necessary to measure full success without intervention, robustness across different objects and robots, safety when people are present, downtime, compute cost and the ability to explain failures.
The announced industrial deployments will also need to be monitored. A one-off demonstration and continuous production have very different requirements. The most instructive data will be the data describing operation over several weeks, the number of human interventions, the modifications required and the conditions under which the system must be shut down.
S1 thus represents a credible advance in in-context robot learning, but not yet proof of a universal robot that learns instantly. Its real impact will depend on its ability to turn a promising demonstration into a reliable, verifiable and responsible process.
Further reading
To place S1 in the context of evolving robotics models, embedded AI and industrial careers, explore these other analyses from the aivancity blog.
Sources
[1] Skild AI, August 2026. Introducing S1: In-Context Learning for Robotics. Accessed September 22, 2026. View the source
[2] NVIDIA, September 10, 2026. Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video. View the source
[3] Robotics and Automation News, September 14, 2026. Skild AI Unveils S1 Robot Foundation Model That Learns Tasks From Video Demonstrations. Read the article
[4] European Commission. Cadre réglementaire européen sur l’intelligence artificielle. Accessed September 22, 2026. View the regulatory framework
[5] European Union, 2023. Règlement (UE) 2023/1230 relatif aux machines. View the regulation