S1 is a robot manipulation foundation model developed by Skild AI and introduced on August 18, 2026 in the technical post "Introducing S1: In-Context Learning for Robotics". Its distinguishing design point is the form of the prompt: a task is specified by a video demonstration held in the model's context rather than by a language instruction, and the model is reported to execute it with no fine-tuning and no post-training. The company's claim is that S1 is the first robotics foundation model to show in-context learning on tasks up to ten minutes long that were never seen during pre-training.
All published figures are Skild's own, measured on two internal benchmark suites. As of August 26, 2026 no independent evaluation, third-party replication, or public benchmark entry had been reported, and the model has no published system card or safety case.
Architecture and inputs
The model takes a video demonstration as the task specification, recorded either as egocentric human video or as a robot rollout, alongside the robot's own observations. Skild states that the demonstration may come from a different scene, viewpoint, or embodiment from the deployment setting, so the policy must infer the demonstrator's intent, functional correspondences, and task progress in order to predict actions at all (Introducing S1: In-Context Learning for Robotics). Parameter count, context length, action rate, and model family are not disclosed. S1 is described as trained on Nvidia AI infrastructure (Introducing S1: In-Context Learning for Robotics).
Training
Pre-training runs on episodic data in which the task is specified only through an in-context demonstration. Skild describes this in meta-learning terms: pre-training is an outer loop that teaches the policy how to learn from context, and at inference the demonstration drives an inner loop without changing any weights. Task diversity and scale are given as the drivers of the capability, on the argument that in a high-diversity regime scene ambiguity can only be resolved by attending to the context. One set of weights is stated to have produced every example in the release post (Introducing S1: In-Context Learning for Robotics).
The training corpus draws on four collection modes rated against hardware proximity, diversity, and scalability — robot teleoperation, UMI hand-held collection, egocentric video, and simulation — which the company states it scales in house rather than betting on one. It reports spending three dollars on quality control for every dollar spent on collection, screening each data point for low-level precision, task coherence, and annotation fidelity (Introducing S1: In-Context Learning for Robotics).
The post announcing S1 states that it is the first in a series and that the training method will be described in later instalments; no training compute, corpus size, or hardware count has been published.
Capabilities and reported results
The central result is a controlled comparison between demonstration prompting and the language prompting used by conventional vision-language-action policies, with both models trained on identical data, architecture apart from the prompt embedding, and compute, across pre-training sets from 1,000 to 100,000 hours.
| Regime | In-context demonstration | Language-prompted |
|---|---|---|
| Seen (in-distribution) tasks, 100k hours | 96% | 89% |
| Unseen (out-of-distribution) tasks, 100k hours | 66% | 9% |
| Seen tasks, 1k hours | 43% | 53% |
All figures are average per-step success rates, measured under a protocol that uses human intervention to recover from failures so that every step is graded; they are not end-to-end task success rates. Skild describes the unseen-task gap as a sevenfold improvement, a ratio of its own two measurements rather than a separately measured quantity. The reversal at 1,000 hours is attributed to in-context models overfitting to a small amount of high-dimensional conditioning data, where language compresses the task specification into a few tokens (Introducing S1: In-Context Learning for Robotics).
Four long-horizon tasks absent from pre-training — plant potting, pancake cooking, pour-over coffee making, and kit assembly — are reported as running up to ten minutes across dozens of manipulation steps from one demonstration. For the plant-potting task the company logs eleven minutes between recording the demonstration and autonomous execution, with most of the surrounding time spent setting up the scene (Introducing S1: In-Context Learning for Robotics).
Robustness
Robustness is measured along two axes — distance from pre-training conditions and distance from the demonstration in context — across five displacement levels running from identical placement (L1) to a placement requiring half the robot's actions to be executed with the opposite arm (L5).
| Level | Language-prompted (shift from training) | S1 (shift from training) | S1 (shift from the prompt) |
|---|---|---|---|
| L1 | 89 | 96 | 96 |
| L2 | 82 | 91 | 93 |
| L3 | 82 | 86 | 89 |
| L4 | 68 | 86 | 75 |
| L5 | 33 | 75 | 46 |
The language-prompted policy is reported to degrade up to three times as much as S1 at L5. S1's robustness to prompt mismatch holds through substituted objects (L4) and falls only where the demonstration implies a substantially different execution plan (Introducing S1: In-Context Learning for Robotics).
Behaviours reported as emergent
Skild reports that S1 completed tasks despite objects being slid away or swapped mid-execution and the lighting changed, none of it shown in the prompt; that it retries after a failure rather than proceeding, including on out-of-distribution tasks such as assembling a skateboard wheel; that it substitutes a cup when the prompt uses a watering can and only a cup is available, and tops off a nearly full glass where the prompt fills an empty one; and that it corrects flawed demonstrations, performing with a controlled motion a step in which the demonstrator dropped an egg. The company reads the last of these as the model treating the demonstration as a specification of the goal rather than a trajectory to reproduce (Introducing S1: In-Context Learning for Robotics).
Relation to fine-tuning
S1's advantage over weight updates is confined to the low-data regime. Skild puts one in-context demonstration at roughly 380 post-training episodes, stating that the crossing point was estimated by interpolating between measured points, and reports that a post-trained vision-language-action policy passes S1 with enough data, reaching 86% at 2,000 demonstrations against S1's 66% from a single demonstration (Introducing S1: In-Context Learning for Robotics).
Lineage
Skild's earlier published model is LocoFormer, an in-context learner for locomotion that adapts by accumulating live experience in its prompt, published as arXiv:2509.23745 by Min Liu, Deepak Pathak and Ananye Agarwal in September 2025. The company's published timeline runs LocoFormer in September 2025, first in-domain in-context manipulation results in February 2026, a first pancake flip in May 2026, and the S1 release in August 2026. Skild separately markets a general robot control layer as the Skild Brain; its published material does not state whether S1 is that system, a component of it, or a separate manipulation-specific model (Introducing S1: In-Context Learning for Robotics).
Availability
No API, weights, pricing, or general availability had been announced as of August 26, 2026. The release is a technical post rather than a product launch. Skild states that S1 is at work with commercial partners, without naming them or describing the scale of deployment (Introducing S1: In-Context Learning for Robotics).
Comparison with peer models
The closest published comparison is GEN-1.5 from Generalist AI, announced the following day, which reports acquiring new manipulation tasks from a 3–12 second demonstration held in a 30-second context window with no gradient updates, averaging 59% across ten short-horizon tasks (GEN-1.5: Embodied Foundation Models are One-Shot Learners). Skild's post names the concurrent work and argues it is "constrained to short horizons, in-distribution tasks, or both"; Generalist's own account does not accept that characterization. The two are not directly comparable — Generalist reports average task success across manipulation primitives, Skild average per-step success under human-intervention recovery on internal suites — and no common benchmark places them on one scale.
Other embodied foundation models pursue generality along different axes: Helix from Figure AI, Gemini Robotics 2 from Google DeepMind, and the embodiment-flexible models of Physical Intelligence. S1's published claim concerns task horizon and novelty from a single demonstration; the others' concern task breadth and instruction following.
Relationships
- developer: Skild AI
- instance-of: In-Context Learning — the model is built around demonstration conditioning rather than fine-tuning.
- instance-of: AI Robotics, Foundation Models
- depends-on: Scaling Laws — the company frames the capability as a product of task diversity and pre-training scale.
- related: GEN-1.5, Generalist — the concurrent in-context claim S1's release argues against.
- related: Helix (Figure AI Vision-Language-Action model), Gemini Robotics 2, Figure AI, Physical Intelligence — peer embodied foundation-model efforts.
- related: Emergent capabilities, Embodied Perception (Tacit Knowledge vs. Pattern Recognition)
Sources
- Introducing S1: In-Context Learning for Robotics — Skild AI, "Introducing S1: In-Context Learning for Robotics," August 18, 2026.