S1
AI model · Robotics / world model by Skild AI
The in-context learner for general-purpose robotics.
S1 is a flagship robotic foundation model that enables robots to learn new tasks by watching a single video demonstration. It is an omni-bodied system designed to control any robot morphology in unstructured, real-world environments.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Launched
- 2026-08-18
- Languages
- English
Category: Machine Learning
- Modalities
- Video → Robot actuation
- Sizes and variants
- S1 (Flagship model)
- Released
- 2026-08-18
- Model Type
- Robotic Foundation Model
- Learning Architecture
- In-context learner
- Input Modality
- Video demonstrations
- In-context learner: learns from a single video demonstration
- Omni-bodied: controls any robot for any task
- Bridges the embodiment gap between humans and robots
- Fine-tunes for new skills with <1 hour of robot data
Overview
S1 is Skild AI's flagship robotic foundation model designed to bring general-purpose intelligence to physical hardware. Built as an in-context learner, it allows robots to acquire new skills by watching a single video demonstration, effectively overcoming the traditional data bottleneck of manual teleoperation. The model is omni-bodied, meaning it provides a unified brain capable of controlling diverse robot types, from industrial arms to quadrupeds, across unstructured real-world environments. By leveraging internet-scale human video data, S1 bridges the embodiment gap between human movement and robotic actuation. It is intended for industrial and commercial applications, enabling rapid deployment of autonomous skills with less than one hour of robot-specific fine-tuning data.
Learning and Adaptation
- In-Context Learning: The model is built from the ground up as an in-context learner, allowing it to acquire new skills by watching a single video demonstration.
- Cross-Embodiment Mapping: It bridges the embodiment gap by mapping human actions from video to robot-specific actuation across different morphologies like industrial arms or quadrupeds.
- Low-Data Fine-tuning: The model can be fine-tuned to perform new skills using video demonstrations and less than one hour of robot-specific data.
- Observational Learning from Video: It leverages internet-scale human video data, including egocentric headcam footage and instructional videos, to overcome the robotics data bottleneck.
Robotic Control and Interaction
- Omni-bodied Control: The model is designed as a unified, omni-bodied brain capable of controlling any robot type regardless of its physical configuration.
- Physical Intelligence: It possesses a foundation of kinematics and dynamics that allows it to internalize intent from visual demonstrations and map actions to its own body.
- Unstructured Environment Navigation: The model enables robots to navigate and perform tasks in unstructured, messy, and dangerous real-world environments.
- Low-level Skill Execution: It can execute low-level skills such as grasping, handover, and navigation, which are abstracted away for application developers.
Use cases
- Industrial operators show a robot a single video of a new task to have it learn the skill autonomously.
- Warehouse managers deploy the model across different robot types to automate repetitive packing and sorting tasks.
- Robotics engineers use the model to bridge the gap between human demonstrations and robot actuation with minimal data.
- Security teams implement the model to allow inspection robots to navigate dangerous or unstructured environments.
Specifications
- Modalities
- Video → Robot actuation
- Domains
- General-purpose robotics, industrial automation, warehouses, security and inspection
- Sizes and variants
- S1 (Flagship model)
- Fine-tuning
- Supported using video demonstrations and <1 hour of robot data
- Released
- 2026-08-18
- Model Type
- Robotic Foundation Model
- Learning Architecture
- In-context learner
- Input Modality
- Video demonstrations
- Fine-tuning Data Requirement
- < 1 hour of robot data
- Supported Morphologies
- Omni-bodied (Multi-robot)
- Primary Use Case
- General-purpose robotic control
How it works
- Foundation Training: The model is trained on internet-scale human video data to understand physical kinematics and dynamics.
- In-Context Demonstration: A user provides a single video demonstration of a task being performed.
- Cross-Embodiment Mapping: The model internalizes the intent from the video and maps the actions to the specific robot's morphology.
- Rapid Fine-tuning: The model is fine-tuned with less than one hour of robot-specific data to master the new skill in the real world.
FAQ
How does S1 learn new tasks?
S1 is an in-context learner that can acquire new skills by watching a single video demonstration of a human performing a task.
What types of robots can S1 control?
The model is omni-bodied, meaning it can control various robot morphologies, including 7-DOF industrial arms, quadrupeds, and mobile manipulation platforms.
How does the model handle the difference between human and robot bodies?
Skild AI bridges this gap by fine-tuning the model to map human visual demonstrations to robot-specific forces and torques using less than one hour of robot data.
What kind of data was used to train S1?
S1 uses internet-scale human video data, such as YouTube instructional videos and first-person headcam footage, to learn kinematics and dynamics at scale.
Added by AntSpace from public sources (skild.ai).