Skip to content

SBP-Guided MPC to Overcome Local Minima in Trajectory Planning

Authors: Emily Hannigan, Bing Song, Gagan Khandate, Ji Yin, Maximilian Haas Heger, Matei Ciocarlie

Published: 2019 (Workshop Paper)

Source: IEEE International Conference on Robotics and Automation (ICRA)

Algorithm: SBP-Guided MPC

Summary

Uses an RRT solution trajectory to warm-start iLQR, preventing the optimizer from getting stuck in local minima and simultaneously refining the crude RRT path into a smoother, more optimal trajectory.

Abstract

Trajectory planning is often a difficult task for highdimensional systems, especially those with non-linear dynamics. Two common methods for trajectory planning in non-linear systems are Model Predictive Control (MPC) and kinodynamic Sampling-Based Planning (SBP). In this paper, we focus on a variant of MPC called the iterative Linear Quadratic Gaussian (iLQG) algorithm[1]. iLQG has been shown to be fast and effective for generating trajectories to reach a goal. However, when optimizing for non-linear dynamics, it runs the risk of falling into local minima. Unlike iLQG, SBP methods such as Rapidly-exploring Random Trees (RRT) [2] can be robust to local minima because they explore by taking random actions instead of following a gradient. However, for similar reasons, SBP often produces inefficient trajectories. We combine these two algorithms to take advantage of the specific strengths of each. We show that by using an RRT-produced trajectory as a warm start for iLQG, we can overcome local minima while still producing an efficient trajectory. On a specific system model (a robot snake), we show that the combination of these two algorithms allows the robot to reach a goal faster than the use of either algorithm alone in most cases.

Tags

  • Trajectory planning

  • Trajectory optimization

  • MPC

  • iLQR

  • RRT

  • Sampling-based planning

  • Local minima

  • Hybrid planning