Keywords: MPC, variational inference, kernel methods
TL;DR: We propose task-space kernels for measuring similarity between control sequences in a task relevant space that incorporates spatio-temporal info. We use these kernels to promote diversity in a particle-based MPC framework.
Abstract: Model predictive control (MPC) has a proven track record for delivering robust performance in many challenging control tasks, however non-linear system dynamics and non-convex costs can make these problems challenging to solve. By taking a probabilistic view and using approximate inference to solve the optimal control problem, we can improve the exploration of the search space. We use a non-parametric approximate inference method that finds diverse solutions using a kernel function. We propose Task-space Kernels which define similarity between solutions in task-relevant spaces and better capture spatio-temporal information. This helps us generate smooth, diverse, and low-cost trajectories for complex robotic problems. We demonstrate this strategy empirically on two multi-modal manipulation tasks with a 7DoF robot where our state, and end-effector space kernels achieve lower average costs and steps over the baselines.
Submission Number: 6
Loading