Key takeaways
- Skild AI introduced S1 in August 2026, and NVIDIA described the companies’ development and deployment collaboration on September 10.
- S1 uses a demonstration video as runtime context instead of updating its weights for every new task.
- On Skild’s internal unseen long-horizon benchmark, cumulative per-step success was 66% for video-conditioned S1 and 9% for a language-conditioned policy.
- The 66% figure is not full autonomous task completion: humans recovered the state after failures so later steps could still be graded.
- Fast task setup must be evaluated separately from repeatability, recovery, cycle time and independent safety controls.
One video becomes the runtime task prompt
Imagine that soil, a plant, a pot and a watering can have just arrived at a robotics lab. A person records one repotting demonstration, and 11 minutes later a robot starts attempting the task from that video. This is the scene highlighted by S1, a robot foundation model that Skild AI introduced in August 2026. NVIDIA described S1 and the companies’ broader physical-AI collaboration on September 10. The results discussed here come from company publications; they are not yet an independently reproduced public benchmark or an open-model evaluation.
A conventional deployment often starts with a person teleoperating the robot for many demonstrations. Engineers then fine-tune a task-specific policy and validate it under deployment conditions. S1 is designed to reduce that final round of task-specific training. A video prompt conveys the intent, objects and sequence, while one pretrained model maps that context onto the robot body and workcell that are currently in front of it.
This is called in-context learning. The analogy is a language model that receives a few examples in its prompt and answers a new question without changing its weights. It does not mean that the robot is trained from scratch on one video. S1 already depends on broad pretraining at scales up to 100,000 hours in the experiments described by Skild; the new video specifies a task on top of that prior experience.
Video specifies motion that language can omit
Video can specify details that language leaves ambiguous. “Make pour-over coffee” does not say how to seat the filter, how quickly to tilt the kettle or in what order to use the tools. A demonstration includes motion, timing and object interaction. Skild says its pretraining setup forces the policy to infer intent, functional correspondence and task progress even when the demonstrator, camera viewpoint and robot embodiment differ.
Skild reports that S1 performed previously unseen tasks such as plant potting, pancake cooking, pour-over coffee and kit assembly for as long as ten minutes. A long task is harder than a single pick because the robot must compose dozens of steps, track what has already happened and notice mistakes. One small failure can change the state for every step that follows.
What the 66% versus 9% benchmark measured
The headline comparison is 66% versus 9%. Skild trained a video-conditioned policy and a language-conditioned VLA with the same data, nearly identical architectures and the same compute. At 100,000 hours of pretraining, the video-conditioned approach reached 66% cumulative per-step success on unseen tasks, while the language-conditioned baseline reached 9%. This is an internal benchmark; it does not establish a sevenfold gain for every robot and task.
The denominator matters. Skild says the benchmark tasks lasted four to eight minutes and the score was the average cumulative per-step success across tasks. To grade later steps, a human intervened after failures to restore the rollout state, mainly for the language baseline. The 66% number therefore must not be rewritten as “the robot autonomously completed the entire task 66% of the time.”
The conditions behind the “380 demonstrations” claim
Skild also estimates that one in-context demonstration matched roughly 380 task-specific post-training demonstrations. The crossing point was interpolated between measured data points and applied to unseen tasks lasting four to ten minutes. The company says collecting those 380 teleoperated demonstrations took 50–100 hours. Yet task-specific post-training eventually surpassed the one-shot result: with 2,000 demonstrations it reached 86%, compared with S1’s 66%. Rapid setup and peak reliability are different objectives.
A factory demo is not production reliability
The factory implication is substantial. If a line can adopt a changed product or process without another long data-collection and fine-tuning cycle, changeover cost could fall sharply. Skild says it is working with Foxconn and NVIDIA on dual-arm assembly of Blackwell systems, including busbar and limit-block installation and fastening 16 screws. A demonstrated workflow, however, is not the same as validated long-duration yield, cycle time or uptime.
One successful video clip reveals almost nothing about reliability. A system with a low success rate can still produce an impressive clip after enough attempts. Before deployment, customers need repeated-trial completion rates, intervention counts, recovery rates, mean and tail cycle times, contact-force and damage data, and results under shifted lighting, objects and tool placement.
Safety remains a separate layer. Understanding a demonstrated task does not ensure that the robot will avoid people or limit contact forces. Real installations still require independent speed and force limits, collision detection, emergency stops, workspace controls and risk assessment. In-context learning may become a fast way to specify what the robot should do; it is not a substitute for a safety-rated control system.
Four questions for real deployment
The most important idea in S1 is the attempt to move robot teaching from a new dataset and training run toward a demonstration and runtime context. Four questions will determine whether that shift reaches production: How many minutes does task setup truly take? How many consecutive trials succeed? How far can the robot recover without help? What is the total cost including interventions and validation? Those answers—not one polished demo—will show whether one-video prompting becomes factory productivity.
Sources reviewed
- Skild AI: Introducing S1 — In-Context Learning for Robotics (August 2026)
- NVIDIA: Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video (September 10, 2026)
- Skild AI: The Hidden Pillar of Robotics (September 10, 2026)
Related reading
Sean Woo — I have spent more than 15 years shaping robotics technology and business direction. Drawing on public technical documents, research papers and company announcements, I explain changes in robotics and AI in accessible terms. The interpretations in this publication do not represent the official position of any company or organization.