Skild AI

A model across different robot bodies

Skild AI develops robot policies rather than a single humanoid design. Its architecture combines a slower policy for manipulation and navigation with faster control that converts commands into joint positions and motor torques. Training across different bodies is intended to make experience transferable between machines. [1]

Its July 2025 public research account shared demonstrations and results from 2024. The company described pretraining with simulation and human video, followed by targeted robot data. That publication documented a trained research system; it did not provide downloadable weights or establish that every robot could run it. [1]

Specifying behavior through a demonstration

The later S1 account explores using a video as the task prompt. Skild reports execution of new manipulation sequences without task-specific weight updates, distinguishing in-context adaptation from fine-tuning. Its examples and evaluations remain developer-reported results, with task selection and experimental conditions that matter when assessing generality. [2]