Event
Ph.D. Research Proposal: Chak Lam Shek
Tuesday, November 25, 2025
1:30 p.m.
https://umd.zoom.us/j/7140250593?pwd=OTVma003K3hzTG9zeUhiN0FUS0RGQT09&omn=98900325652&jst=2
Sarah Pham
301 473 2449
spham124@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Chak Lam Shek
Committee:
Professor Professor Pratap Tokekar (Chair)
Professor Dinesh Manocha
Professor Kaiqing Zhang
Date/time: Tuesday, November 25 from 1:30pm to 3:30pm
Location: https://umd.zoom.us/
Title: From One to Many: Reinforcement Agents with Uncertainty-Aware Generalization
Abstract:
Artificial Intelligence (AI) systems are increasingly expected to operate in complex and uncertain environments, where robustness to practical limitations is essential. Recent advances in agent-based AI demonstrate the potential of small, resource-constrained agents that cooperate to achieve tasks beyond the capacity of any individual agent. Specifically, this line of research contributes to the development of autonomous systems for applications such as autonomous driving, aerial robotics, and humanoid robots, where agents must perceive, reason, and act safely in dynamic real-world settings. However, this shift introduces fundamental challenges that demand principled algorithmic solutions. This dissertation is centered on reinforcement learning (RL), developing methods that extend RL frameworks to explicitly address three classes of constraints—limited capabilities, resource constraints, and partial observability—enabling both single-agent and multi-agent systems to operate robustly and efficiently under uncertainty.
First, we address resource constraints, where efficiency and reliability must be carefully balanced under costly or limited operations. For single-agent navigation, we focus on the problem of deciding when to localize in scenarios where frequent observations are expensive and continuous localization is infeasible. We introduce a reinforcement learning framework that learns selective localization policies, minimizing resource usage while ensuring that the probability of failure remains bounded. In multi-agent systems, we investigate how to allocate constrained update budgets more effectively across heterogeneous agents. By adapting thresholds based on the relative contribution of each agent, we enable faster convergence, more efficient use of resources, and more stable coordination than uniform allocation schemes.
Second, we consider agents with limited sensing capabilities, arising either from the physical design constraints of the robot or from limitations in perception algorithms. Specifically, we study the locomotion problem for a single agent operating in unstructured terrains and the coordination problem of sensor deployment for multiple energy-constrained robots. To address the locomotion challenge, we design a context-aware framework that leverages language-based contextual cues alongside sensor data to guide navigation policies under limited perception. For the sensor deployment task, we develop an estimation method that accounts for uncertainty in airdropped sensor locations and introduce a submodular optimization framework to plan energy-constrained unmanned aerial vehicle (UAV) routes that maximize information gain despite stochastic landing outcomes.
Third, we tackle partial observability, where agents must plan and act with incomplete or noisy information. In the single-agent case, we develop a hierarchical framework that uses language-driven subgoal generation, reusable options, and action-level policies to improve exploration efficiency and enhance generalization across tasks. In the multi-agent case, we exploit structural patterns of locality within decentralized systems. By partitioning agents into strongly related groups using dependency structures, we achieve more accurate credit assignment, reduce the curse of dimensionality, and improve cooperative learning efficiency compared to global-reward methods.
