Learning Warmstart: Accelerating Trajectory Optimization with Self-Attention Siamese Network
The computational complexity of trajectory optimization poses significant challenges for real-time robotics applications, where timely solutions are critical. This thesis introduces a novel approach to accelerate trajectory optimization by learning to warmstart the optimization process using a Self-Attention Siamese Network. The proposed method maintains a library of previously solved trajectory optimization problems and their solutions, then employs a neural network to identify the most similar past problem when faced with a new optimization task. The solution to this similar problem serves as an initial guess for the new problem, potentially reducing solve time.
The approach is evaluated on two dynamic systems: an inverted pendulum and a quadrotor. Results demonstrate that appropriate warmstarts can significantly improve optimization outcomes on the pendulum, particularly for torque-constrained problems. For quadrotor trajectory optimization, the Self-Attention Siamese Network consistently outperforms both random selection and Euclidean distance-based methods, achieving up to 20% reduction in solve times. These findings emphasize the importance of constraint-awareness when selecting initial guesses from library and demonstrate the potential of neural network approaches to accelerate optimization for robotic systems.
Advisor: Professor Constantinos Chamzas (WPI)
Committee: Professor Mahdi Agheli (WPI) and Professor Guanrui Li (WPI)
Zoom: wpi.zoom.us/j/6363139422