Deep Learning based Robotic Motion Generation for Manipulating Granular and Viscous Materials
This research project is in competition for funding with one or more projects available across the EPSRC Doctoral Training Partnership (DTP). Usually the projects which receive the best applicants will be awarded the funding. Find out more information about the DTP and how to apply.
Start date: 1 October 2019
This project aims to investigate in developing Deep Learning based methods for robots to learn to manipulate such materials using tools, such as scoops, shovels, or trowels, and generate motions autonomously in real-time.
The Deep Learning algorithms will primarily include two methods, namely Learning from Demonstration (LfD) and Reinforcement Learning (RL), that will build a model to map robot primitive motions, such as spreading, scooping, or reaching, to the deformation of materials observed from sensors like 3D cameras and force sensors.
Granular and viscous media are ubiquitous in our daily life, ranging from food like dough or beans to construction materials like concrete, soils, or sands. However, research in autonomous robotic manipulations of these materials is at its infancy. There are only a limited number of attempts in real-time motion planning to manipulate granular or viscous media using robots.
One main challenge is the material deformation, due to the viscosity, granularity, and viscoelasticity. Traditional robotic motion planning is not scalable in unstructured environments due to explicitly designed models. To autonomously plan motions in real time, viscous or granular characteristics need to be considered. However, physics-based numerical modelling is usually considerably computational and impractical for real-world environments.
Project aims and methods
Active exploration with Reinforcement Learning requires extremely large datasets for training, and will be very time-consuming to obtain enough data through continuous trial and error operations. The project will start with a supervised manner by using LfD for robots to learn basic skills from human demonstrators. LfD allows robots to master new skills from observations of a human’s demonstration that can be tele-operated or by vision observation. RL will continue from the above learned model and explore optimal actions in a limited search space. Ultimately, the robot will be able to perform similar tasks through autonomous motion generation.
The following are proposed:
- Developing LfD algorithms to allow robots to learn basic skills in simple scenarios. The experiments will be carried out with the Kuka iiwa robot in the robotics laboratory. The initial datasets will be obtained.
- Researching on 3D deformation and mapping to robot motion segments.
- Developing RL algorithms by designing reward functions as metrics to evaluate performance of corresponding actions and states and policies for optimal action selection.
- Evaluating performance thoroughly and experimentally to benchmark its generalisation capability.
- Writing up PhD thesis.