
Introduction
Skild AI introduced S1 in August 2026 as an in-context robotics foundation model that takes a video demonstration as the task prompt and attempts unseen tasks without task-specific weight updates.
Reported examples include plant potting, pancake making, pour-over coffee and kit assembly. Skild says some tasks last up to 10 minutes and span dozens of manipulation steps; these are company-reported results and should be evaluated with complete-task success and intervention data.
Source review date recorded for this article September 27, 2026.
In-context robot learning cuts the setup loop
A conventional deployment may collect many demonstrations then retrain a policy. S1 uses the demonstration as context at run time.
Skild reports that one plant-potting test moved from recording the example to autonomous hardware execution in 11 minutes.
Long tasks expose compounding failure
NVIDIA reports about 66 percent success per step in Skild tests on new multistep tasks. A long sequence can fail when any one step fails.
Sequence-level completion rate, recovery count and intervention time would give a fuller view of performance over ten-minute tasks.
Video demonstrations can hide contact information
A single camera view may miss force, depth or hand contact. The model has to infer those details or recover using robot sensors during execution.
Tests with occlusion, moved objects and different viewpoints can show how far one-video learning reaches before another example is needed.
S1 is now also a base for post-training research
On September 23, Skild published a separate physical self-play result that uses S1 as a strong base model for reinforcement-learning post-training in simulation. That work should not be merged with the single-video in-context learning claim because it adds a different training stage.
The distinction is useful for readers: one-video prompting tests whether S1 can infer a task from context without task-specific weight updates, while self-play tests how the base model can improve through additional reinforcement learning.
One example video can leave the target object's contact state ambiguous
A video of a person assembling a kit may show the order of operations but hide how tightly a connector was pressed or whether a part clicked into place. An in-context model must infer some of those conditions or use its robot sensors during execution. A successful short demonstration does not establish that a new robot can reproduce the same contact mechanics without calibration, physical feedback or intervention.
To evaluate S1 on a new task, use one recorded example, declare whether any additional instructions or rehearsal data were provided and run the robot on held-out objects. Count completed full sequences, incomplete steps and manual resets. Do not combine the separate physical self-play reinforcement-learning work with the claim of no task-specific weight updates during one-video prompting; those are different experimental mechanisms.
Conclusion
The strongest follow-up metric is complete-task success across many unseen multistep tasks with recovery and intervention counts included.