Robotics is increasingly automating real-world tasks,from manufacturing and packaging to minimally invasive surgery. Robots could also inspect infrastructure in environments too hazardous or inaccessible for humans, such as tunnels, dams, pipelines, railways, and power plants.

Despite the potential for safer environmental assessments, human agents still conduct most inspections. Computer scientists are developing computational models to plan robot trajectories for inspecting environments and executing desired actions.

Researchers at Purdue University and LightSpeed Studios have developed a new training-free computational technique for generating inspection plans from written descriptions,guiding robots as they inspect specific environments. Their approach, detailed in a paper published on the arXiv preprint server, uses vision-language models (VLMs) to process both images and written texts.

“Our paper was inspired by real-world challenges in automated inspection, where generating task-specific inspection routes efficiently is critical.”

The VLM takes a text description and a 3D environmental map as input. It then outputs smooth trajectories that conform to the user’s text description, including targets, orders, and spatial relationships. this allows for personalized inspection plans tailored to specific needs.