Online controller adaptation with meta-learned models Hersh Sanghvi

Format:

Book

Thesis/Dissertation

Author/Creator:

Sanghvi, Hersh, author.

Contributor:

University of Pennsylvania. Computer and Information Science., degree granting institution.

Language:

English

Subjects (All):

Robotics.

Computer science.

Information science.

0771.

0984.

0800.

0723.

Local Subjects:

Robotics.

Computer science.

Information science.

0771.

0984.

0800.

0723.

Genre:

Academic theses

Physical Description:

1 online resource (138 pages)

Contained In:

Dissertations Abstracts International 87-12A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Owing to significant algorithmic improvements in recent years, robots are moving out of controlled, confined spaces and into the real world. Tuning the control parameters of these systems to operate in a wide range of uncertain environments and novel tasks is a significant challenge. Often, there is no single set of parameters that works in every environment or on every task, necessitating adaptation to different operating conditions. However, existing methods fall short in some key ways. Classical methods for controller adaptation and tuning are restrictive in their assumptions, while modern data-driven methods struggle with sample efficiency and out-of-distribution generalization. This thesis presents data-driven methods for online adaptation of controller parameters that leverage prior knowledge while remaining robust to domain shifts. By learning predictive models that map controller parameters to task-relevant performance measures, the proposed methods enable fast controller adaptation without explicit system identification. Two complementary approaches are presented. Online Continuous Controller Adaptation with Meta-Learned Models (OCCAM) introduces a meta-learning framework for training uncertainty-aware performance models that are rapidly adapted to online data via recursive updates. Meanwhile, Trajectory-Aware Controller Optimization (TACO) formulates online controller tuning as a trajectory-conditioned optimization problem, enabling task-aware gain selection for various maneuvers. These methods enable fast adaptation in the presence of domain shift. OCCAM and TACO are evaluated on a broad range of robotic platforms and tasks, including simulated racing cars, quadrupeds, and both simulated and physical quadrotor platforms. Experimental results demonstrate improved adaptation performance compared to classical tuning methods and existing learning-based baselines. We conclude by outlining directions for future work, including initial results on tuning high-dimensional deep-learning-based policies and broader classes of robotic systems

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: A.

Advisors: Taylor, Camillo Jose Committee members: Jayaraman, Dinesh; Matni, Nikolai; Kumar, Vijay; Malik, Jitendra

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247979739

Access Restriction:

Restricted for use by site license