Spring 2025 GRASP SFI: Student Lightning Talks, Session 1

This will be a hybrid event with in-person attendance in Levine 307 and virtual attendance on Zoom.

ABSTRACT

The Spring 2025 GRASP SFI Student Lightning Talks will highlight the research of three GRASP Lab Master’s or early PhD students whose presentation topics have been nominated by their faculty advisors and voted on by their GRASP peers.

Ho Jin Choi (PhD, MEAM)

Layered Perception Architecture: Integrating Multimodal Sensing for Robot Control

Robotic manipulation in unstructured environments presents challenges due to dynamic objects and incomplete state estimation. Inspired by biological vision systems, Layered Perception Architecture (LPA) is introduced as a framework that integrates multi-scale perception, dynamic state estimation, and structured perception-action coupling to improve manipulation capabilities. This talk will discuss LPA’s key components such as single-object shape and pose estimation, their implementation, and preliminary experiments.

Siming He (MSE, ROBO & UG, CIS)

Active Perception for Robust Information Gathering

Active perception enables autonomous agents to gather information efficiently in uncertain environments. I will present two approaches that improve active perception using information-theoretic and game-theoretic principles. The first formulates active perception using a NeRF-based representation to maximize predictive information gain for exploration. The second addresses estimation errors in information gain, developing an online algorithm with sublinear regret. These methods provide insights into improving robustness and efficiency in active perception tasks.

Pei-An Hsieh (MSE, ROBO)

Learning-Based Model Predictive Control for Tight Formation Flight of Quadrotors

Enabling quadrotors to fly in tight formations presents significant challenges due to complex aerodynamic disturbances, particularly downwash effects. In this talk, I will present KNODE-DW MPC, a novel learning-based Model Predictive Control (MPC) framework that integrates first-principle physics models with knowledge-based neural ordinary differential equations (KNODEs) to accurately capture quadrotor downwash dynamics. This hybrid approach achieves high sample efficiency and enables precise trajectory tracking. Simulation and experimental results demonstrate a 40% improvement in trajectory tracking over nominal MPC, culminating in the first successful demonstration of stacked quadrotor flight with only 12 cm of vertical separation. By leveraging this framework, we enhance learning-based control for multi-robot aerial systems, paving the way for safer and more reliable collaborative autonomous operations.