Skild AI has built a robot foundation model called S1 that can watch a person perform a task once, then try that same job on a robot without updating the model’s weights. The company demonstrated the approach with plant potting and pancake flipping, among other tasks that span up to ten minutes and dozens of individual steps.
Robot Demonstrations
According to Skild’s technical report, S1 successfully attempted plant potting, pancake preparation, coffee preparation, and kit assembly—tasks that were not part of the model’s pretraining. The robot’s weights remained unchanged throughout. In the plant potting demonstration, human filming began at 9:16 p.m., and autonomous execution started at 9:27 p.m., accounting for the eleven-minute turnaround highlighted in The Rundown newsletter. Setup commenced at 8:54 p.m., meaning the interval measures only the gap between recording and execution, not the full preparation time.
Controlled Comparison Results
Skild ran a controlled comparison at 100,000 hours of pretraining: demonstration prompting achieved a 66 percent success rate, while language prompting reached just 9 percent. That figure requires scrutiny. It represents an average of cumulative success across task steps and incorporates human recovery interventions. This methodology makes it difficult to determine how often S1 could complete an entire job without any assistance.
The company estimates that replicating S1’s single-video performance through conventional training would require roughly 380 teleoperated demonstrations—equivalent to 50 to 100 hours of operator time. This crossover point was interpolated between measured data points. With 2,000 demonstrations, the additional-training method climbed to 86 percent success, surpassing the single-video result on Skild’s metric.
Practical Significance for Robot Deployment
The practical significance lies in what S1 could change about robot deployment timelines. Teaching a robot a new multi-step job traditionally demands tens to hundreds of hours of task-specific data. Skild’s experiment puts a concrete number on that burden: one video prompt delivered performance the company estimates would otherwise need approximately 380 teleoperated demonstrations. For teams integrating robots into new workflows, that could shift the economics of initial trials.
Collecting repeated demonstrations requires operator time before a team can assess results. If a single recorded example provides a useful starting point, teams might spend less time building an initial dataset and move faster to evaluating actual robot behavior. Skild’s findings support this possibility while leaving real-world deployment savings unmeasured.
The heavy training investment shifts earlier rather than disappearing. Skild’s comparison involved 100,000 hours of pretraining, and a new demonstration still enters the system as a prompt. The efficiency gain comes from applying an existing foundation to another job with minimal additional task data. How broadly this transfers across unfamiliar environments and different hardware remains an open question.
Comparison with Generalist
Generalist offers a parallel data point. In its August 19 announcement of GEN-1.5, the company reported 59 percent average success across ten simple, short tasks after receiving three to twelve seconds of a single demonstration, with no weight updates required. Generalist’s standard “physical prompts” include sensor data and action trajectories, with separate demonstrations of transfer from bare-hand human examples. These results test rapid adaptation in a different way than Skild’s extended tasks.
Both companies’ reports suggest further training still has a role. Generalist reached 83 percent success after ten gradient steps with five minutes of task data, while Skild’s larger demonstration collection outperformed its single-video baseline. A practical workflow could involve demonstrating a task, assessing performance, then collecting targeted data where reliability lags.
Full-Job Testing Importance
This makes full-job testing essential. Skild’s step-based metric, which counts recovery assistance, leaves unresolved how frequently S1 completes an entire task unassisted. Before assigning a workflow to S1, teams would need to measure complete runs, intervention frequency, and recovery behavior under the conditions where the robot will operate.