Recent advances in robotics rely on large datasets of robot trajectories, images and language descriptions. These are used to train transformer-based models, such as vision-language-action models, for common tasks like picking and placing objects. The models are trained mainly on everyday objects and tasks, with datasets of hundreds of thousands of episodes.

In fields such as healthcare and manufacturing, data of this kind is scarce. This project studies how robot skills learned from large, general datasets can be transferred efficiently to such domains, so that robots can take on new tasks from only a small amount of task-specific data.

The project is a doctoral project carried out by Hashim Ismail, supervised by Volker Krueger and Maj Stenmark, within the NextG2Com competence centre.