Machine learning for robotic grasping and manipulation of everyday objects

Robotic grasping and manipulation skills are essential in many robotic applications. However, despite decades of research, robust autonomous grasping and manipulation at a level approaching human skills remain an elusive goal. One main difficulty lies in dealing with the inevitable uncertainties in how a robot perceives the world, based on sensory measurements that can be noisy and incomplete, which poses challenges to avoid failures. Major scientific challenges prevent us from fully integrating robots in our daily lives. Traditional industry robots heavily depend on hard automation that requires pre-specified fixtures and time-consuming programming performed by experienced software engineers. This makes the system not reusable for a new task without a human expert re-programming in detail every movement of the robot. 

This project will focus on developing a robotic system capable of grasping given objects in realistic settings. It considers uncertainties, and lack of specific prior knowledge about the objects. The goal is to enable robots to improve their skills over time continuously. The key idea is to enable them to gradually gain more information about the variety of physical and geometric properties of objects and changes in the environment, using rich multi-modal sensory information from vision and touch, encode knowledge from experience, and build goal-oriented behaviours with dexterity. There will be a focus on a generalisable approach for grasping built from experience and flexibility to include multi-modal data, minimising data requirements within a complete perception, planning, execution and plan correction pipeline.

Area

Machine Learning

UCL leads

• Dr Yasemin Bekiroglu, Department of Computer Science