Research Project

LENA: Lifelong navigation learning using human-robot interaction

Type

National Project

Start Date

01/09/2023

End Date

31/08/2026

Project Code

PID2022-142039NA-I00

Project illustration

Staff

Project Description

Project PID2022-142039NA-I00 funded by MCIN/ AEI /10.13039/501100011033 and by "ERDF A way of making Europe"

Most existing AI algorithms consist of architectures specialized through data-driven tuning (training) processes. Even though these training datasets are chosen to include the maximum number of topologies, once trained, the resulting networks mostly become static. To soften this constraint for the case of robotic autonomous navigation, where a robot can encounter unforeseen situations, the LENA project aims at developing robust and flexible lifelong learning algorithms, developing methods for incremental and online learning, as well as transfer learning. The proposed research will also leverage human-robot interactions and study how learning efficiency can be increased by modeling human-like and reciprocal communication.

Notice the relevance of the LENA research direction in terms of continual learning, as the methods will be able to:

  1. Adapt to changes in data availability over time, making them a valuable tool in applications where data is constantly changing, missing, limited, unbalanced, non-stationary, or new classes are added.

  2. Improve the robot’s understanding and capabilities in the physical world by gathering new experiences and updating its knowledge, which is of special interest in applications where the environment is dynamic and new tasks, sensor modalities, objects, non-stationary data, or multiple tasks might be added.

  3. Have a better understanding of the context of the information robots are processing, thus improving their accuracy in predictions and decisions.

  4. Learn from human-generated data overtime can enhance the ability of AI systems to generate new and innovative ideas.

  5. Better understand and respond to human needs and preferences, leading to more natural and effective interactions.

  6. Learn from human feedback and adjust their behavior over time, reducing the chances of bias being introduced due to changes in the original data distribution or the environment.

Further, all developed techniques will be tested on real robotic navigation data (e.g., noisy or scarce data captured by a real robot).