FLAIRGEORGIA TECH

FLAIR · project

Variational Inference Motion Planning

Motion planning usually seeks one trajectory that reaches a goal while avoiding obstacles. VIMP instead optimizes a distribution over trajectories, providing a way to trade off trajectory cost against stochasticity within a Gaussian variational inference framework.

Planning as probabilistic inference

The method approximates a desired trajectory distribution with a tractable Gaussian. This is equivalent to adding entropy regularization to a motion-planning objective. Changing the strength of that regularization connects stochastic planning with its deterministic counterpart.

The optimization uses natural gradient descent. It also exploits the sparsity induced by factorized trajectory costs, making the computation more scalable than treating every trajectory variable as densely coupled.

Smooth paths and robustness

Experiments with several simulated robot systems examine collision avoidance and trajectory smoothness. They also compare robustness with deterministic baselines in challenging environments. The work shows how optimizing a trajectory distribution can complement conventional path optimization, while retaining a structured numerical formulation.

Publication

A Gaussian Variational Inference Approach to Motion Planning

Hongzhe Yu and Yongxin Chen

IEEE Robotics and Automation Letters · 2023