Symmetries infused safe & scalable multi-robot policies Nikolaos Bousias

Format:

Book

Thesis/Dissertation

Author/Creator:

Bousias, Nikolaos, author.

Contributor:

University of Pennsylvania. Electrical and Systems Engineering., degree granting institution.

Language:

English

Subjects (All):

Robotics.

Engineering.

Computer science.

0771.

0800.

0537.

0984.

Local Subjects:

Robotics.

Engineering.

Computer science.

0771.

0800.

0537.

0984.

Genre:

Academic theses

Physical Description:

1 online resource (160 pages)

Contained In:

Dissertations Abstracts International 87-12B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Deep learning has enabled multi-robot systems to acquire complex, coordinated behaviors that were previously out of reach - yet the black-box nature of learned policies makes it difficult to ensure that they respect the underlying physics and geometry of the systems they control, let alone scale reliably or provide safety guarantees. This dissertation argues that symmetry offers a principled answer to this challenge. Behavioral and geometric symmetries manifest as equivalences among state-action pairs under group transformations: a policy need only learn one mapping per equivalence class, rather than redundantly re-learning the same behavior across all related instances. Embedding such structure as an inductive bias into learned policies, safety certificates and uncertainty quantification methods, reduces effective problem complexity and improves efficiency and generalization - setting symmetry exploitation as a foundational principle for safe and scalable multi-robot intelligence. The first part of this dissertation studies symmetry-aware decentralized policy learning for multi-robot active information acquisition. Exploiting the interchangeability of homogeneous robots, the dissertation introduces a permutation-equivariant graph neural architecture that enables experience sharing across agents and communication topologies. Trained via imitation learning from a centralized expert, the resulting policy scales to larger teams and hidden-state dimensions, generalizes to previously unseen configurations, and remains robust under communication failures and dynamic environments. Next, it extends symmetry exploitation to multi-agent reinforcement learning, with a keen focus on systems with partial or broken intrinsic symmetries. It formalizes conditions under which optimal policies are equivariant and proposes a methodology for embedding extrinsic symmetries when exact system symmetries are insufficient. To parametrize the symmetry-enhanced policy, this dissertation proposes a modular equivariant graph-based architecture for distributed swarming tasks. The symmetries-enhanced policies demonstrate improved learning efficiency, transferability, and scalability on symmetry-breaking quadrotor systems. The following chapter, then, considers safety-critical multi-robot control through distributed Control Barrier Functions. The dissertation shows that safety certificates and safe policies often inherit geometric symmetries from the underlying task and safety specifications, and introduces symmetry-infused multi-agent control barrier functions that improve safety generalization and reduce conservatism while preserving forward-invariance guarantees. Experiments demonstrate improved scalability, transfer to denser and larger swarms, and stronger safety-performance tradeoffs relative to prior learned safe-control approaches. Finally, the dissertation extends symmetry exploitation to uncertainty quantification through equivariant Conformal Prediction, a post-hoc method that symmetrizes arbitrary pretrained predictors. The thesis establishes that this procedure yields sharper conformal prediction sets in expectation while maintaining formal coverage guarantees, and empirically demonstrates reduced uncertainty in long-horizon trajectory prediction tasks. Ultimately, this dissertation positions symmetry not merely as a modeling convenience, but as a foundational principle for safe and scalable multi-robot intelligence. Across decentralized decision-making, safety certification, and uncertainty quantification, it shows how symmetry-aware methods can reduce effective problem complexity, improve transfer and robustness, and maintain formal guarantees in increasingly large and complex multi-robot systems

Notes:

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

Advisors: Pappas, George J. Committee members: Prorok, Amanda; Ribeiro, Alejandro; Matni, Nikolai

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247972785

Access Restriction:

Restricted for use by site license