Learning based motion planning
Title: |
Learning based motion planning |
DNr: |
Berzelius-2023-25 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
Bernhard Wullt <bernhard.wullt@it.uu.se> |
Affiliation: |
Uppsala universitet |
Duration: |
2023-02-27 – 2023-09-01 |
Classification: |
10201 |
Keywords: |
|
Abstract
We want to apply machine learning to solve motion planning problems for manipulators in time-varying environment.
Our approach is to use imitation learning, meaning that we want to mimic an expert motion planner that knows how to solve the problem. The policy and work-space encoding network is parameterized by a deep neural network, hence requires GPU:s for training in a scalable way.
We have previously evaluated our approach in a 2D environments, but now we aim to increase the difficulty as a next step and train our networks for a 3D environment, which increases the required amount of compute needed to train these models.
Once we have a planner that can solve our simulated world, we want to deploy this trained planner in a real world to evaluate its performance.