Skild AI S1 uses NVIDIA physical-AI infrastructure to learn unfamiliar long-horizon robot tasks from one video
The model uses a video demonstration as an in-context prompt without task-specific retraining, while production safety still requires independent validation.
Key takeaways
NVIDIA described Skild AI's S1 robot foundation model on September 10. An operator records a task video and uses it as a prompt; S1 interprets the demonstrated intent, objects, and sequence, then maps them into actions for the robot without updating model weights or performing task-specific post-training. NVIDIA says S1 can handle unfamiliar tasks lasting up to ten minutes, including plant potting, pancake making, pour-over coffee, and kit assembly. In tests reported by Skild, the model succeeded at about 66 percent of steps on new multistep tasks versus 9 percent for a similar system. The companies use NVIDIA infrastructure for synthetic data, training, simulation, reinforcement learning, and deployment. These are vendor-reported demonstrations and tests, not evidence that every factory can operate without supervision; deployment still requires staged safety, reliability, and recovery validation.
Direct answer
NVIDIA described Skild AI's S1 robot foundation model on September 10. An operator records a task video and uses it as a prompt; S1 interprets the demonstrated intent, objects, and sequence, then maps them into actions for the robot without updating model weights or performing task-specific post-training. NVIDIA says S1 can handle unfamiliar tasks lasting up to ten minutes, including plant potting, pancake making, pour-over coffee, and kit assembly. In tests reported by Skild, the model succeeded at about 66 percent of steps on new multistep tasks versus 9 percent for a similar system. The companies use NVIDIA infrastructure for synthetic data, training, simulation, reinforcement learning, and deployment. These are vendor-reported demonstrations and tests, not evidence that every factory can operate without supervision; deployment still requires staged safety, reliability, and recovery validation.
Verified facts
S1 takes an operator-recorded video as a prompt, interprets intent, objects, and sequence, and maps them into robot actions without updating weights or task-specific post-training.
NVIDIA says S1 handles unfamiliar tasks lasting up to ten minutes, including plant potting, pancakes, pour-over coffee, and kit assembly. A potting demo went from recording to autonomous hardware execution in eleven minutes.
In Skild-reported tests on new multistep tasks, S1 achieved about 66 percent success at each step versus 9 percent for a similar system; the post does not present these figures as cross-environment safety certification.
What changed
- One operator video can provide task context instead of starting with a dedicated dataset
- Long-horizon behavior composes learned skills and attempts recovery after disturbances
- Synthetic data, simulation, reinforcement learning, and inference optimization form one development chain
- Vendor demonstrations must become independent, station-by-station and step-by-step acceptance tests
Impact for AI users
A single video demonstration can reduce the upfront data and programming work needed to adapt a robot to a changing process. Yet errors compound across a multistep task, so a reported 66 percent per-step success rate does not translate directly into an acceptable end-to-end yield. Teams should treat demonstration learning as a way to shorten configuration, while retaining safety cages where appropriate, force and collision limits, human emergency stops, and a defined safe state after failure.
Operating checklist
- Choose one reversible task in a low-speed, low-payload isolated cell with no human entry, and version the demonstration video.
- Cover object shifts, occlusion, grasp failure, tool faults, and emergency stops in simulation, recording both per-step and end-to-end success.
- During hardware tests, limit speed, force, workspace, and continuous runtime, with an independently triggered human stop.
- Expand stations and task scope only after safety thresholds, recovery paths, and end-to-end yield all pass.
AI frontier news and analysis, AI software and model tools, AI skill tutorials and validation methods, and AI account and permission guidance
FAQ
Does S1 require retraining for every new task?
Skild says no task-specific post-training or weight update is required; the video demonstration supplies task context.
Does 66 percent per-step success mean a ten-minute workflow succeeds 66 percent of the time?
No. The post reports per-step success in Skild's tests; end-to-end success across many steps must be measured separately.
Can one-video learning replace human safety controls?
No. The article describes capabilities and recovery examples, but it does not replace site risk assessment, physical safeguards, emergency stops, or acceptance testing.
Summary
S1 shows a path for shortening robot task setup with video context. Production readiness still depends on measurable safety at every stage, acceptable end-to-end yield, and predictable recovery after failure.
This AI-assisted article is checked by ENHE AI automation for official sources, bilingual fields, media rights, page safety, and historical duplication before publication.
What this means for everyday users
A single video demonstration can reduce the upfront data and programming work needed to adapt a robot to a changing process. Yet errors compound across a multistep task, so a reported 66 percent per-step success rate does not translate directly into an acceptable end-to-end yield. Teams should treat demonstration learning as a way to shorten configuration, while retaining safety cages where appropriate, force and collision limits, human emergency stops, and a defined safe state after failure.
Tools you may use
Related tutorials
Related Tools And Tutorials
Use the following ENHE AI sections to continue from the news signal into tool selection, account-service guidance, or practical learning.
Related reading
AWS benchmarks OpenAI models by cost per correct outcome instead of token price alone
AWS published an open-source benchmarking approach on September 11 for comparing GPT-5.6 Luna, Terra, and Sol on Amazon Bedrock with GPT-5.4 mini and nano on the OpenAI API. Instead of ranking models by dollars per million tokens, the harness measures cost per correct answer, cost per passing multi-turn research result, and cost per rubric-accepted professional deliverable. It uses one Responses API client and records timestamped result files, frozen grading prompts, and chart inputs. The post also documents important limits: sample sizes range from 48 to 198 items, Bedrock models ran with reasoning disabled, and API baselines used their defaults. AWS describes this as a practical configuration comparison rather than an estimate of intrinsic capability. Teams should rerun the method on 50 to 100 of their own tasks and include failures, retries, turns, latency, review, and rework in the decision.
A production release gate for Structured Outputs: validate the schema, refusals, incomplete responses, and semantics
Structured Outputs lets compatible models return data that adheres to a supported JSON Schema, providing a stronger contract than JSON mode, which guarantees valid JSON but not schema adherence. OpenAI documents two forms: function calling for connecting a model to application tools, and a structured response format for shaping user-facing output. In strict schemas, object fields must be listed as required and objects must set additionalProperties to false; an optional value can be represented with a union that includes null. Schema conformance covers the shape of a successful response, not factual accuracy, business rules, or a completed request. A production release gate should pin the model and schema version, test normal output, refusals, output-token truncation, content-filter incompleteness, and unsupported schema errors, then run independent semantic validation before any write or external action.
Inside OpenAI Habitat: scaling online storage past 70M requests per second and migrating from Python to Rust
OpenAI published an engineering account of Habitat on September 11. The company says its online storage platform now handles more than 70 million requests per second, supports products used by over one billion people each week across almost 40 regions, and serves more than 500 petabytes of data. Habitat began as a Python client library for GPTs in 2023, then became a standalone service so deployments, observability, access controls, audit logging, and platform improvements could be managed centrally. In the second quarter of 2026, two engineers used Codex and GPT-5.5 to rewrite the service in Rust. OpenAI reports that the Rust service now handles 95% of production requests and is six times more CPU-efficient and fifteen times more memory-efficient than the Python version. These are internal production measurements, so other teams should adopt the migration discipline rather than assume identical gains.
Salesforce introduces Slackforce Surfaces for AI-built dashboards, reports, and interactive work inside Slack
Salesforce introduced Slackforce Surfaces on September 11, allowing people to ask Slackbot to turn conversations, Salesforce records, Slack history, and connected enterprise data into dashboards, reports, executive HTML decks, calculators, and other interactive interfaces. A Surface can be pinned to a channel so a team can filter, explore, discuss, and act from one shared workspace. Salesforce says Surfaces operate under existing permissions, limiting users to information they are already allowed to see. The product is available for customers on Enterprise+, Business+, Pro, Legacy, and Free Teams when Slackbot is enabled. However, the company separately states that the live-data feature begins rolling out in October. Teams should therefore verify workspace eligibility, refresh behavior, record-level grounding, and downstream actions rather than assume every demonstrated live capability is active immediately.
Three Google-supported Android XR projects premiere at Venice with Gemini conversations and spatial filmmaking
Google highlighted three projects supported by its 100 ZEROS filmmaking initiative on September 11. The projects premiered during the 83rd Venice International Film Festival and are designed for Android XR. NEVATARS, a mixed-reality animated short directed by Andy Serkis, includes Gemini-powered conversations and interactive moments. Galápagos: The Last Eden is an immersive journey narrated by Margot Robbie that lets viewers interact with a Gemini-powered digital Charles Darwin. Sedona, a psychological-thriller teaser from Asylm Studios, uses Google's 2D-to-3D XR auto-spatialization technology for a YouTube experience. Google says the broader lineup uses conversational AI powered by Gemini and Vertex AI, spatial film, and stereoscopic depth while keeping human creativity at the center. The announcement presents creative experiments rather than evidence that AI tools replace writers, directors, performers, or audience testing.
NVIDIA and Palantir turn supply-chain allocation expertise into a governed Nemotron training loop
NVIDIA published a supply-chain case study with Palantir Foundry on September 10. The workflow combines a governed Ontology, cuOpt optimization, planner decisions and rationales, point-in-time backtesting, and post-training of Nemotron 3.5 Lightning for material allocation recommendations. NVIDIA reports that its post-trained 30B model reached 86.7% allocation-decision accuracy on the development benchmark, compared with 55.5% for Nemotron 3 Ultra and 17.5% for the base Lightning model. The company also says a human planner reviews recommendations and makes the final call, while accepted, edited, and overridden outcomes feed future governed retraining. This is an official case study and development benchmark for a bounded allocation task. It does not establish broader general intelligence or general superiority for the 30B model beyond the specialized data, task, and evaluation design.
