Event
Ph.D. Dissertation Defense: Levi Burner
Friday, September 12, 2025
9:30 a.m.
IRB 4105
Souad Nejjar
301 405 8135
snejjar@umd.edu
ANNOUNCEMENT: Ph.D. Dissertation Defense
Name: Levi Burner
Committee:
Professor Yiannis Aloimonos (Chair)
Dr. Cornelia Fermuller
Professor Perinkulam Krishnaprasad
Professor Dinesh Manocha
Professor Guido de Croon
Professor Nikhil Chopra (Dean's Representative)
Date/time: Friday, September 12, 2025 at 9:30 AM
Location: IRB 4105
Title: Fundamentals of Embodied Representation: Robotics Without a Ruler
Abstract:
Imagine sitting at your desk, looking at various objects on it. While you do not know their exact distances from your eye in meters, you can reach out and touch them. Instead of an externally defined unit, your sense of distance is inherently tied to your action's effect on your embodiment. Animals ranging from insects to humans use such a concept of distance to determine what actions to take. They must because most animals do not know what an external scale, such as the meter, is. Instead, they must make do with an internal unit, or embodied unit, that is somehow measured using the signals available from the body. In contrast, today's robots almost exclusively rely on the meter. Their bodies are measured in meters, their sensors are calibrated to the meter, and consequently, their control, planning, and vision systems use the meter. Further, extensive effort is put into calibrating these systems end ensuring those calibrations do not degrade over time, else the robot will stop working.
If robots could represent the world using their own sense of distance, they would not require such calibration or precise engineering. Instead, they could use uncalibrated sensors, with uncalibrated bodies, to accomplish tasks such as clearing obstacles, jumping gaps, and manipulating objects. Inspired by this problem, this dissertation develops a visuomotor approach through which robots can accomplish such tasks without prior knowledge of an external unit. The key to the approach is using a system's own actions, or control inputs, as internal feedback from which a unit is implied and used to estimate quantities such as the body's size, motor dynamics, or position of objects in the world. The resulting techniques are called ``Embodied Representation'' because they consist of measurements in terms of the signals available to the robot's body itself, without calibration to an external scale.
The development of Embodied Representation is detailed in this dissertation, resulting in techniques and algorithms for fundamental problems encountered by embodied systems. First, a specific method for using time-to-contact, a bio-inspired visual representation, with acceleration, is used to achieve stable closed-loop control even when the units are embodied and thus unknown. Subsequently, a general framework ``Embodied Visuomotor Representation'' is developed for estimating and using such representations. The resulting algorithms for uncalibrated clearing and jumping mirror natural behaviors observed in bees and gerbils. Next, the use of internal feedback for manipulation is studied, specifically for key insertion. The robot compares its own wiggling signal to tactile feedback in order to guide an insertion process. The result can insert keys into four types of locks. Furthermore, in an assembly benchmark, it outperforms a reinforcement learning baseline trained on the objects. Finally, the impact of vibration on visual perception, as induced by limbed or winged locomotion, is studied. This results in a specialized video stabilization algorithm, which represents the world in a coordinate frame different from the system's body and stabilizes video from a tailless ornithopter known for its aggressive shaking.
Name: Levi Burner
Committee:
Professor Yiannis Aloimonos (Chair)
Dr. Cornelia Fermuller
Professor Perinkulam Krishnaprasad
Professor Dinesh Manocha
Professor Guido de Croon
Professor Nikhil Chopra (Dean's Representative)
Date/time: Friday, September 12, 2025 at 9:30 AM
Location: IRB 4105
Title: Fundamentals of Embodied Representation: Robotics Without a Ruler
Abstract:
Imagine sitting at your desk, looking at various objects on it. While you do not know their exact distances from your eye in meters, you can reach out and touch them. Instead of an externally defined unit, your sense of distance is inherently tied to your action's effect on your embodiment. Animals ranging from insects to humans use such a concept of distance to determine what actions to take. They must because most animals do not know what an external scale, such as the meter, is. Instead, they must make do with an internal unit, or embodied unit, that is somehow measured using the signals available from the body. In contrast, today's robots almost exclusively rely on the meter. Their bodies are measured in meters, their sensors are calibrated to the meter, and consequently, their control, planning, and vision systems use the meter. Further, extensive effort is put into calibrating these systems end ensuring those calibrations do not degrade over time, else the robot will stop working.
If robots could represent the world using their own sense of distance, they would not require such calibration or precise engineering. Instead, they could use uncalibrated sensors, with uncalibrated bodies, to accomplish tasks such as clearing obstacles, jumping gaps, and manipulating objects. Inspired by this problem, this dissertation develops a visuomotor approach through which robots can accomplish such tasks without prior knowledge of an external unit. The key to the approach is using a system's own actions, or control inputs, as internal feedback from which a unit is implied and used to estimate quantities such as the body's size, motor dynamics, or position of objects in the world. The resulting techniques are called ``Embodied Representation'' because they consist of measurements in terms of the signals available to the robot's body itself, without calibration to an external scale.
The development of Embodied Representation is detailed in this dissertation, resulting in techniques and algorithms for fundamental problems encountered by embodied systems. First, a specific method for using time-to-contact, a bio-inspired visual representation, with acceleration, is used to achieve stable closed-loop control even when the units are embodied and thus unknown. Subsequently, a general framework ``Embodied Visuomotor Representation'' is developed for estimating and using such representations. The resulting algorithms for uncalibrated clearing and jumping mirror natural behaviors observed in bees and gerbils. Next, the use of internal feedback for manipulation is studied, specifically for key insertion. The robot compares its own wiggling signal to tactile feedback in order to guide an insertion process. The result can insert keys into four types of locks. Furthermore, in an assembly benchmark, it outperforms a reinforcement learning baseline trained on the objects. Finally, the impact of vibration on visual perception, as induced by limbed or winged locomotion, is studied. This results in a specialized video stabilization algorithm, which represents the world in a coordinate frame different from the system's body and stabilizes video from a tailless ornithopter known for its aggressive shaking.
