Doctoral Thesis Oral Defense - Arjun Lakshmipathy
May 4, 2026 12:00PM—2:00PM
Location:
Newell-Simon 1305 and Zoom
Speaker:
ARJUN LAKSHMIPATHY ,
Ph.D. Candidate, Computer Science Department, Carnegie Mellon University
https://www.andrew.cmu.edu/user/aslakshm/
Humans use their hands to effortlessly manipulate objects of arbitrarily complex geometries and physical properties every day; however, adapting these behaviors to dexterous robots and virtual characters is a difficult task. Understanding how humans exploit contact to perform these manipulations has the potential to greatly advance progress towards this goal.
Unsurprisingly, research efforts have analyzed contact in the context of dexterous manipulation for decades. We now have numerous metrics for evaluating grasp quality in terms of contacts, sophisticated models of contact states, efficient means of computing physically simulated contacts, and strategies that exploit contact correspondences between hands and objects to synthesize grasps and manipulations. But the majority of existing works fundamentally characterize contact the same way: as points, lines, or planes of interaction.
But contact in the real world is much more complicated. Real bodies instead interface with one another via areas of contact which greatly vary with the geometries of the contacting surfaces. If we wish to model the complexities of manipulations as they actually occur, then we must progress beyond such simplifying assumptions and deal with the messy nature of reality.
This thesis aims to do so by presenting foundational frameworks and algorithms for the modeling, capture, mutation, and exploitation of contact areas. Our intention is to establish the foundations necessary to elevate contact regions to first-class primitives and demonstrate their inherent value across a range of practical applications in dexterous manipulation and adjacent domains.
First, we introduce three novel contact area models alongside operations supported by each model designed to run on real geometries rather than primitive shapes. Next, using area-based primitives, we introduce: a set of intuitive artist tools for digitally drafting high quality grasps, a kinematic motion retargeting pipeline for dexterous manipulations, a contact-driven control framework for multi-fingered hands in physical simulation, and two practical extensions to different domains. We then shift our focus to the real world by introducing approaches for capturing and reconstructing contact areas during human-object and human-human interactions. Finally, we present an end-to-end system architecture framework for constructing fully functional robot systems from contact-rich human demonstrations.
Thesis Committee:
Nancy S. Pollard (Chair)
Jessica K. Hodgins
Keenan Crane
Zackory Erickson
C. Karen Liu (Stanford University)
In-person and Zoom
Contact
Matt Stewart