Georgia Institute of Technology
Efficient Iterative Proximal Variational Inference Motion Planning for Nonlinear Systems
Abstract
dc:description.abstractMotion planning under uncertainty can be cast as a stochastic optimal control problem, where the optimal trajectory corresponds to the solution of a maximum a posteriori (MAP) estimation. However, computing the posterior distribution is often intractable in practice. To approximate it, we formulate an optimization problem in the space of Gaussian distributions, framing it as a variational inference (VI) problem in the path distribution space. To tackle this challenge, we propose Parallel KL-Proximal Gaussian Variational Inference Motion Planning (P-GVIMP), a scalable and robust variational inference framework for trajectory optimization. Our approach leverages the KL-Proximal Point algorithm to achieve stable updates and fast convergence. By exploiting the sparse structure of the motion planning factor graph, we factorize the collision cost and apply Gaussian Belief Propagation (GBP) for efficient marginal inference. To further accelerate inference, we employ sparse Gauss–Hermite quadrature for expectation estimation, enabling GPU-parallelized computation across dense trajectories. Thanks to this parallel implementation, the method scales efficiently to high-dimensional systems such as 7-DOF manipulators. To extend this framework to nonlinear stochastic dynamics, we incorporate Statistical Linear Regression (SLR) to iteratively construct time-varying linear approximations of the dynamics. At each iteration, the linearized system is solved using the proposed P-GVIMP algorithm as a subroutine. This results in a principled iterative method—i-PGVIMP—that enables robust trajectory inference even in complex nonlinear environments. We validate our approach on various robotic systems, including 2D/3D point robots, 7-DOF manipulators, and nonlinear planar quadrotors. The results demonstrate the proposed framework’s robustness, efficiency, and scalability in both linear and nonlinear dynamics. Compared with baselines such as GVIMP, our method achieves significantly faster inference through parallel computation while maintaining reliable trajectory planning under uncertainty.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Chang, Zinuo
- Advisors dc:contributor.advisor
-
- Chen, Yongxin
- Vela, Patricio A.
- Committee members dc:contributor.committeemember
-
- Zhao, Ye
- Chou, Glen
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81504