CLEAR-BOT: CoLlaborative intElligence for human-Aligned RoBOTs

We are experiencing rapid advances in robot learning driven by foundation models and human-in-the-loop machine learning methods, but aligning the robot-learned representations to the goal-oriented nature of human representations is still a major limitation that can lead to undesirable and unsafe actions.

For example, a robot vacuum may learn to repeatedly lay out dust and re-clean, misunderstanding the true human objective of a clean floor.

In this project, we plan to develop new methods for creating robots that learn from human input without overly querying humans but remain aligned with the human intent behind the feedback signal.

We propose a hybrid framework where baseline models are iteratively refined with feedback obtained directly from humans or indirectly (e.g. from foundation models).

Our key contributions include exploring methods for robots to detect their own misalignment and proactively elicit additional feedback, finding an optimal trade-off between direct and indirect evaluative feedback, as well as developing novel benchmarks for this type of hybrid AI systems. A PhD student and the main applicant will work on this project which will run for 48 months.

We aim to develop the next generation of intelligent robots that operate safely around people by improving robot learning while minimizing the need for human feedback.