Robots Atlas>ROBOTS ATLAS
Skild AI S1

Skild AI S1

S1
Skild AI's flagship robotic foundation model and in-context learner: turns a single task video demo into robot actions without fine-tuning.
โœ“ Activeโณ Limited accessRobotics foundation model
Release date
18 August 2026

Overview

Skild AI S1 is the flagship robotic foundation model unveiled on August 18, 2026 by the startup Skild AI. It is built as an in-context learner: it translates a single in-context video demonstration of a desired task into robot actions, without requiring fine-tuning or post-training at deployment. S1 can perform novel, previously unseen manipulation tasks up to about 10 minutes long, and shows robustness to scene perturbations and object substitutions, along with common-sense reasoning and error recovery. Demonstrated tasks include plant potting, pancake cooking, pour-over coffee making, and kit assembly โ€” all from a single video prompt. The model was pre-trained on episodic data where the task is specified only through an in-context demonstration, combining teleoperation data, egocentric video, and simulation. S1 is built on NVIDIA AI infrastructure.

Classification
Robotics foundation model
Access & deployment
Weights: Closed
Key parameters
๐Ÿ“ฅ Input: video, robot sensors, robot state data
Robotics
Dexterous manipulationRobot manipulationEmbodied task planningRobot controlObject affordance understanding

Technical specification

License
Proprietary
Modalities
โฌ‡ Input
videorobot_sensorsrobot_state_data
โฌ† Output
robot_actionsrobot_commandsmotion_trajectoriesmanipulator_control

Capabilities and applications

Native model capabilities
Video Understanding
Category: video
Reasoning
The model's ability to perform multi-step logical inference, solve complex problems and decompose tasks into steps.
Category: reasoning
Robotics
Dexterous manipulationRobot manipulationEmbodied task planningRobot controlObject affordance understanding

Benchmark results

3 benchmarks
Long-horizon seen tasks (4-8 min)
success rate
96%
๐Ÿ“„ Oficjalny blog Skild AI (S1)
Vendor-reported result.
Long-horizon unseen tasks @100k training hours
success rate
66%
๐Ÿ“„ Oficjalny blog Skild AI (S1)
Vendor-reported result.
Demonstration efficiency
sample efficiency
1 demo ~ 380 post-training episodes
๐Ÿ“„ Oficjalny blog Skild AI (S1)
Equivalence of one video demonstration.

Technical architecture