Date: Friday, January 23, 2026

Time: 9:00AM - 10:30AM ET

Location: Online via Microsoft Teams

Virtual:  Link; Meeting ID: 282 710 312 091 6 ; Passcode: 49JM9FT9 

Tomohiro Sasaki

Ph.D. Robotics Student

School of Aerospace Engineering

Georgia Institute of Technology

Committee:

Dr. Koki Ho

School of Aerospace Engineering

Georgia Institute of Technology

·  Dr. E. Glenn Lightsey

School of Aerospace Engineering

Georgia Institute of Technology

· 

·  Dr. Lu Gan

School of Aerospace Engineering

Georgia Institute of Technology

· 

·  Dr. Bogdan Epureanu

Department of Mechanical Engineering

University of Michigan

· 

·  Dr. Naoya Ozaki

·  Institute of Space and Astronautical Science

·  Japan Aerospace Exploration Agency

· 

Abstract:

Robotic autonomy in safety-critical settings demands methods that are both reliable and adaptive. Spacecraft rendezvous, quadrotors in wind, and mobile robots in changing environments all require decisions under constraints, uncertainty, and shifting objectives. Model-based optimal control provides physics-grounded predictions and hard constraint satisfaction, while reinforcement learning adapts from experience but can violate constraints and lacks interpretability. This work develops a unified framework for constrained trajectory optimization that combines these paradigms, targeting real-time performance with safety guarantees. The methodology is grounded in interior-point iLQR/DDP and introduces three main advancements. First, a multiple-shooting formulation integrated with interior-point methods enables robust convergence from dynamically infeasible initializations without sacrificing efficient Riccati structure. Second, chance-constrained DDP provides probabilistic safety margins using closed-loop covariance propagation, and GPU acceleration enables richer uncertainty evaluation to support hardware validation. Third, implicit KKT differentiation yields a differentiable constrained MPC layer that integrates with broader autonomy stacks and preserves hard constraints. Validation spans representative platforms, including spacecraft, aerial vehicles, and ground robots. Overall, the work enables learning-driven adaptation while maintaining real-time performance and hard constraint satisfaction.