Abstract: Techniques for determining a text description of a scene in an environment by a computing device are discussed herein. The scene can represent a simulated scene for use in a simulation between a vehicle (e.g., an autonomous vehicle) and one or more objects. The computing device can implement a machine learned model to generate a text description of the simulated scene based on receiving a title and one or more scenes as input. Text descriptions can be validated and used for testing a controller of a vehicle, or other robotic device, during simulation.