Project Description

Project PID2024-162505OA-I00 funded by MCIN/ AEI /10.13039/501100011033 and by ERDF, UE

Over the last years, the improvements in certain fields of AI have raised questions, both in research and media about when robots will be present in our homes, performing our daily tasks. However, there are still several challenges to address before this can become effective, aside from hardware costs. Large companies and institutions are devoting a substantial amount of funds in trying to learn generalist robot policies that make robots capable of learning and adapting to a large variety of circumstances that happen in everyday life, with a very limited amount of new real-world information, if not one -shot. However, the current state of the art focuses on high-level orders, such as planning sequences of actions, and then generating motions with a policy deciding the next displacements of the robots. These trendy approaches, for which a lot of data is being gathered, ignore a key aspect of humanrobot interaction. Low level aspects such as the robots strength, effort, or speed are often left out of the scope, and the tasks are executed with default low-level motion planners and controllers.

This project aims at providing low level controllers and motion planners with information coming from the environment, the humans verbal commands and preferences, and prior experience. All these elements are to be connected to a common feature space, a latent space within a suitable mathematical manifold, which contains numerical variables, from which a mapping to an impedance controller is to be learned. This will result in a variable rigidity of the robot manipulator that accommodates to both the environment and the human. Verbal instructions can be matched with context and learned human preferences to significantly bias the robots behavior to accomplish a given task and take a step closer to a natural, general pHRI. Within this framework, we aim at deploying this newer technology as a proof of concept in a real robotic scenario.