Title: Dynamics Model Learning For Risk-Aware Planning and Control With Perceptual Inputs

Data: December 1, 2025
Time: 12:00 pm - 2:00 pm
Location: CODA 1215 Midtown
Virtual Link: https://gatech.zoom.us/j/94372341191

Jason Gibson
Robotics PhD Student
School of Interactive Computing
Georgia Institute of Technology

Committee:
Dr. Evangelos Theodorou (Advisor) 
School of Aerospace Engineering
Georgia Institute of Technology

Dr. Frank Dellaert 
School of Interactive Computing
Georgia Institute of Technology

Dr. Samuel Coogan
School of Electrical and Computing Engineering
Georgia Institute of Technology

Dr. Lu Gan
School of Aerospace Engineering 
Georgia Institute of Technology

Dr. Shreyas Kousik
School of Mechanical Engineering
Georgia Institute of Technology

Abstract:
The formulation of autonomous navigation problems are becoming increasingly important as we begin to include more autonomous systems in everyday life. These systems need to be able to adapt quickly to novel inputs and handle the continuous and varied nature of challenges in the real world. We want autonomous systems that are able to appropriately represent and plan with uncertainty in action and perception. The system should naturally modulate its own behavior as a function of uncertainty and the environment by seamlessly adapting between aggressive and conservative actions as a human would. Furthermore, the system should be easily configurable along the continuum of risk with minimal tuning parameters. Throughout the document, we will show a framework for uncertainty-aware trajectory optimization achieve these goals.

Our work focuses on appropriate representations of dynamical uncertainty for real-time optimization. We will review prior work on combining well-understood parametric equations with learned representations to get the best of both approaches. These representations have been incorporated into a real-time trajectory optimization approach using MPPI that will demonstrate human-like risk-sensitive behavior that will scale across multiple environments. This behavior will be demonstrated in miles-long experiments with full-sized vehicles in the real world. We will propose future extensions that will improve the environmental context of uncertainty and increase the traversal speed. Semantic information will be extracted from the environment in a transferable and scalable way using a general feature vector from a visual foundation model. Finally, we will propose new approaches to reduce how conservative the approach is through better preemptive approximations of closed-loop uncertainty.