Introducing S1: In-Context Learning for Robotics

Introducing S1: In-Context Learning for Robotics is a technical post published by Skild AI on August 18, 2026 on the company's own blog, credited to the Skild AI Team and self-cited as @article{skild2026s1} (Source: skild.ai). It introduces S1, the company's flagship manipulation model, and argues that the useful product of robot pre-training is in-context learning: a task is specified by a video demonstration placed in the model's context, and the model executes it without fine-tuning or post-training. The post states that it is the first in a series, and that how S1 is trained will be described in later instalments.

The post is a vendor self-report. Both evaluation suites are Skild's own, no result has been independently reproduced, and the figures are reported as average per-step success rates measured under a protocol that uses human intervention to recover from failures so that every step is graded.

Central argument

The post frames robotics against the history of language modelling. Encoder models such as BERT (Devlin et al., 2019) were effective at understanding language but required fresh data collection and fine-tuning for each application; the transition to ChatGPT, Skild argues, was driven by in-context learning, which is "what separates research prototypes like GPT-1 and GPT-2 from current frontier language models." Robotics, on this account, "has thus far been stuck in the BERT era": robustly executing a new task still requires tens to hundreds of hours of demonstration data in deployment conditions and a task-specific training run.

The argument turns on a finding the post cites from Oh et al. (2026): where post-training data is large enough and densely covers the task, policies trained from scratch with no pre-training match the peak performance of a post-trained foundation model. Skild draws the inference that "the main purpose of pre-training — if not the only one — is to enable, for robotics, the same shift we saw from BERT to GPT-3."

Evaluation axes proposed

The post argues that the field's difficulty is not defining in-context learning but evaluating it, and proposes two axes along which any such claim should be read:

• Whether the demonstrated tasks were seen during pre-training (in-distribution) or not (out-of-distribution).
• Whether the tasks are short-horizon and atomic (5–25 seconds) or long-horizon, requiring new skills to be learned or composed in previously unseen ways.

It quotes Jim Fan for the position that once enough data is available many behaviours are zero-shot without fine-tuning, and that whether in-context learning "truly works" depends on how far the test sits from training. Applying its own axes, Skild states that concurrent approaches — naming Generalist AI (2026) and Jiang et al. (2026) — "largely cover tasks which are short-horizon or already present in the pre-training distribution," and claims 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.

Method as described

Pre-training runs on episodic data in which the task is specified only through an in-context demonstration. Because the demonstration may come from a different scene, viewpoint, or embodiment, the policy must infer the demonstrator's intent, functional correspondences, and task progress in order to predict actions at all. In meta-learning terms the post describes pre-training as an outer loop that teaches the policy how to learn from context, with the demonstration driving an inner loop at inference time without changing weights. One set of weights is stated to have produced every example in the post. S1 is described as trained on Nvidia AI infrastructure.

The post sets out a data strategy rather than a single source. It rates four collection modes on hardware proximity, diversity, and scalability, and states that Skild scales all four in house:

The company states that for every dollar spent collecting data it spends three on quality control, screening each data point for low-level precision, task coherence, and annotation fidelity.

Reported results

Scaling comparison against language prompting

The post's principal quantitative result is a controlled study on a filtered subset of the pre-training pool, training an in-context policy and a language-conditioned policy on identical data, architecture (apart from the prompt embedding), and compute, across pre-training sets from 1,000 to 100,000 hours. Evaluation runs on two internal suites, one drawn from the pre-training distribution and one built from unseen tasks.

The ordering reverses at the smallest scale, which Skild attributes to language compressing task specification into a few tokens while in-context models can overfit to a small amount of high-dimensional conditioning data. On unseen tasks the company describes the 66%-against-9% gap as a sevenfold improvement; the figure is a ratio of two of its own measurements rather than a separately measured quantity. Its explanation is that language lacks grounding for a novel action primitive such as flipping a pancake, and is too coarse to specify a novel chaining of known primitives, whereas a demonstration "specifies a mode" and spells out the composition directly.

Long-horizon unseen tasks

Four tasks the model was not trained on — plant potting, pancake cooking, pour-over coffee making, and kit assembly — are reported as running up to ten minutes and spanning dozens of manipulation steps from a single visual demonstration, including behaviours the post describes as novel at test time: digging into soil to make room for a plant, pressing a coffee filter into a funnel, and flipping a pancake.

The post publishes a timeline for the plant-potting run as an operational measure of deployment cost: materials arrived at the office at 8:54 PM, the scene was set up and recording started at 9:16 PM, one egocentric human demonstration was recorded at 9:22 PM, and the robot began executing autonomously at 9:27 PM — eleven minutes from demonstration to autonomous execution, with most of the elapsed time spent moving furniture.

Robustness to distribution shift

Two robustness axes are measured separately: distance from the pre-training conditions, and distance from the demonstration held in context. Five displacement levels are defined — L1 same objects and placement; L2 objects shifted 15 cm and 30°; L3 shifted 30 cm and 45°; L4 objects swapped for different ones with the same affordances plus roughly 15 cm vertical shifts; L5 placement chosen so half the robot's actions must be executed with the opposite arm.

Skild reports that under L5 the language-prompted policy degrades up to three times as much as the in-context one, and reads the pattern as vision-language-action policies being robust to small pose variation but failing once novel motions are required. Robustness to prompt mismatch degrades only at L5, where the demonstration implies a substantially different execution plan.

Demonstration efficiency

The post expresses the unseen-task gap as a demonstration count: a vision-language-action policy was post-trained on unseen tasks with 1 to 2,000 teleoperation samples, and the in-context policy — never post-trained — appears as a horizontal reference at its single-demonstration success rate. Skild puts one in-context demonstration at roughly 380 post-training examples, stating explicitly that the crossing point was estimated by interpolating between measured points. Post-training passes the in-context policy with enough data, reaching 86% at 2,000 demonstrations against the in-context policy's 66%; the company states it expects that gap to shrink as in-context pre-training scales.

Behaviours reported as emergent

The post catalogues qualitative behaviours under the heading "Emergent Properties," each reported as observed rather than trained for:

• Robustness to perturbation. Objects were slid away as the robot approached, swapped mid-execution, and the lighting changed, none of it shown in the prompt; the post states S1 completed the task regardless.
• Mistake recovery. S1 is reported to retry rather than proceed blindly after a failure, including on out-of-distribution tasks such as assembling a skateboard wheel, without the targeted data collection such behaviour usually requires.
• Common-sense substitution. Where the prompt watered a plant with a watering can and only a cup was available, S1 used the cup; where the prompt filled a glass with juice and the glass was nearly full, it topped it off.
• Demonstration correction. In one prompt the demonstrator drops an egg prematurely; S1 is reported to perform the step with a controlled motion instead, which the post reads as the model treating the demonstration as a specification of the goal rather than a trajectory to reproduce.

Lineage and terminology

The post places S1 after 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. A figure gives the intervening timeline: September 2025 LocoFormer, February 2026 first in-domain in-context results, May 2026 "S1 flips first pancake," August 2026 S1 release. The company states S1 is already deployed with commercial partners, without naming them.

The post calls the model S1 throughout. The company's general robotics foundation model is marketed as the Skild Brain and its locomotion work as LocoFormer; the post does not state how the three relate.

Position taken against a concurrent claim

Skild names GEN-1.5 from Generalist AI as a concurrent approach and argues that recent work in manipulation "remains constrained to short horizons, in-distribution tasks, or both." Generalist's own account of the same capability, published the following day, makes overlapping one-shot claims on ten short-horizon manipulation primitives (GEN-1.5: Embodied Foundation Models are One-Shot Learners). Both are vendor self-reports on separate internal benchmarks with different measurement conventions — Generalist reports average task success, Skild average per-step success under human-intervention recovery — and no common benchmark places the two on one scale.

Provenance

Published at skild.ai on August 18, 2026, on the developer's own domain, and surfaced more widely by Techmeme on August 25. The capture is complete in text; figures and embedded videos are represented by their captions and by the numeric values stated in the body, so no chart value beyond those is recoverable. Six references are cited: Brown et al. (2020) on few-shot language learning, Devlin et al. (2019) on BERT, Generalist AI (2026) on GEN-1.5, Jiang et al. (2026) on context scaling for robot policies (arXiv:2607.15275), Liu et al. (2025) on LocoFormer (arXiv:2509.23745), and Oh et al. (2026) on force sensing and policy learning (arXiv:2606.12406).

Relationships

• supports: S1 (Skild AI) — the primary record for the model's reported capabilities.
• instance-of: In-Context Learning — the post is the developer's own statement of the paradigm applied to manipulation.
• instance-of: AI Robotics — a robotics foundation-model capability report.
• contradicts: GEN-1.5: Embodied Foundation Models are One-Shot Learners — names the concurrent approach and argues it is limited to short horizons or in-distribution tasks; the two are competing vendor accounts rather than reconcilable measurements.
• depends-on: Foundation Models, Scaling Laws — the argument rests on what pre-training scale is for.
• related: Skild AI, Generalist, Figure AI, Nvidia & TSMC — AI Compute Infrastructure — the developer, the named rival, a peer effort, and the infrastructure supplier the post credits.
• related: Emergent capabilities — the qualitative behaviours are reported as unplanned products of pre-training.