Robotic vision competition targets overconfident robots

A new competition in robotic vision aims to address the problem of overly-confident robots misidentifying objects in the real world.

The Robotic Vision Challenge, launched by the Australian Centre for Robotic Vision which is headquartered at QUT, invites challengers to detect objects in video data from high-fidelity simulation of three different types of domestic service robots.

The robot competitors must detect objects in cluttered indoor settings, like lounge rooms and kitchens, with the added challenge of encountering day and night scenes.

Professor Peter Corke, director of the centre, said the competition threw down the gauntlet to the world’s robotics and computer vision research communities to join the centre’s mission to develop new robotic vision technologies to expand the capabilities of ‘truly useful’ robots.

“Big global competitions have been very effective in computer vision research, but they haven’t really pushed the envelope for the sorts of problems an actual robot encounters in the real world,” said Corke.

“This new competition aims to solve that. It’s the world’s first robotic vision challenge; not a computer vision challenge.”

Dr Niko Sünderhauf, centre chief investigator, received a $72,000 Google Faculty Research Award to support its development by a team of centre researchers based at QUT.

“Today’s best machine learning methods for object detection are often overly confident in their own knowledge and have no good way of expressing when they don’t really recognise what they see,” said Sünderhauf.

“With this challenge we hope to motivate researchers around the world to develop new probabilistic methods that know when they don’t know.”

The Robotic Vision Challenge, which offers a $5000 cash prize, has been accepted as a key workshop at the world’s largest computer vision forum, the Conference on Computer Vision and Pattern Recognition (CVPR), in June.

The four-day conference takes place in Long Beach, California from June 16-20, 2019.

“This is very exciting because we involved global research communities at the concept stage of the challenge last year, seeking ideas and insights at major forums like CVPR and the Robotics: Science and Systems conference.

“We want to continue that global partnership, bringing robotics and computer vision communities together and encouraging new thinking on problem-solving,” said Sünderhauf.