Thursday, June 18
11:00 am to 12:30 pm
Newell-Simon Hall 4305
11:00 am to 12:30 pm
Newell-Simon Hall 4305
Integrating Learning and Collaboration for Human-Robot Alignment
Abstract: The alignment problem for robots considers how robots can learn to behave in accordance with human values. Today, robot learning paradigms enable humans to provide data (e.g., preference labels or demonstrations), which the robot uses to update its behavior (e.g., reward model or policy) to better align with human intentions. However, the current paradigm requires the human to constantly supervise, provide new feedback, and—more fundamentally—perfectly understand where the robot is misaligned. Even if the robot eventually learns a perfect model of how the human wanted it to behave, the overall human-robot interaction during alignment could have been demanding, confusing, or arduous for the person.
This dissertation argues that alignment should not be viewed solely as a problem of learning the correct task behavior, but also as a problem of supporting humans throughout the alignment process itself. We propose that the learning process should be fundamentally collaborative, in which robots actively participate in alignment through introspection, communication, uncertainty estimation, sharing control, and interaction planning.Our early works investigate how robots can learn and adapt collaborative strategies from interaction, including adapting to partner preferences and proactively communicating robot capabilities to improve coordination. We then introduce model-agnostic uncertainty quantification methods for interactive robot learning using online conformal prediction, enabling robots to calibrate deployment-time uncertainty from intermittent human feedback. Building on these capabilities, we develop collaborative active learning methods in which robots request assistance when uncertain, calibrate intervention policies using human feedback, strategically allocate human effort across multitask learning settings, and negotiate control handoffs through shared-control interaction mechanisms. Finally, we study the relationship between uncertainty and human intervention behavior, showing that robot uncertainty and human judgments of intervention necessity are only partially aligned, motivating methods that explicitly calibrate robot assistance requests to human preferences. The outcomes are robots that know when they do not know, communicate uncertainty, request help strategically, support human teachers during learning, and adapt interaction itself as part of the alignment process.Thesis Committee Members:
This dissertation argues that alignment should not be viewed solely as a problem of learning the correct task behavior, but also as a problem of supporting humans throughout the alignment process itself. We propose that the learning process should be fundamentally collaborative, in which robots actively participate in alignment through introspection, communication, uncertainty estimation, sharing control, and interaction planning.Our early works investigate how robots can learn and adapt collaborative strategies from interaction, including adapting to partner preferences and proactively communicating robot capabilities to improve coordination. We then introduce model-agnostic uncertainty quantification methods for interactive robot learning using online conformal prediction, enabling robots to calibrate deployment-time uncertainty from intermittent human feedback. Building on these capabilities, we develop collaborative active learning methods in which robots request assistance when uncertain, calibrate intervention policies using human feedback, strategically allocate human effort across multitask learning settings, and negotiate control handoffs through shared-control interaction mechanisms. Finally, we study the relationship between uncertainty and human intervention behavior, showing that robot uncertainty and human judgments of intervention necessity are only partially aligned, motivating methods that explicitly calibrate robot assistance requests to human preferences. The outcomes are robots that know when they do not know, communicate uncertainty, request help strategically, support human teachers during learning, and adapt interaction itself as part of the alignment process.Thesis Committee Members:
Henny Admoni (Chair)
Reid Simmons (Chair)
Andrea Bajcsy
Anirudha Majumdar (Princeton)