Jacobs Robotics achieve great success at ICRA Robot Challenge

The Robotics Group from Jacobs University finished second at the 'Solutions in Perception Challenge' held during the IEEE International Conference on Robotics and Automation (ICRA) in Shanghai, China. The team was closley beaten by UC Berkeley. Stanford University came third.

May 12, 2011

Competitors at the 'Solutions in Perception Challenge', which ended on May 12, had to develop software that enables robots to recognize and properly locate objects as basis for manipulation. This had to be done in a variety of situations and also included detecting new objects on the fly.

Eight different approaches were put forward in the competetion, seven solutions from independent institutions and one from California-based software company Willow Garage (as a professional basis for comparison). Together with the US National Institute of Standards and Technology (NIST), Willow Garage had organized the event and also provided one of their very advanced PR2 robots for testing and demonstration.

At the start of the contest the Jacobs Robotics group took the lead, but had to admit defeat at the very end when UC Berkeley came in slightly ahead of them. The Californians reached a final score of 68.78% compared to Jacobs' result of 66.41%. Stanford University, meanwhile, ended up in third place with 53.61%.

The baseline solution from Willow Garage achieved 50.86% and a fourth place, while the remaining other teams got an average score of 11.25%.

The scores reflect the number of correctly recognized objects, the precision of object localization as well as the runtimes of the algorithms. The Jacobs Robotics team had almost exactly the same recognition rate as Berkeley, but clearly outperformed them in precision of object localization.

However, in their computation time Berkeley was almost twice as fast. In the end, it was just a small differences in performance on the final test cases that finalized the decision in favour for Berkeley.

The algorithms for this challenge had to use a Microsoft Kinect camera. This device, which was originally designed to act as motion controller for the Xbox, provides 3D range information as well as a RGB color images. This allows for developing methods that not only take the visual appearance of objects in terms of colors and textures into account, but also their 3D shapes.

Overall, the competetion showed that object recognition and localization are still major challenges for robots and computer programs in general. When objects are placed in cluttered scenes with many occlusions, robots are faced with an even tougher situation.

To handle the data from a low-cost device like a Kinect is also rather challenging. The solution developed at Jacobs Robotics breaks up the classical borders between subsequent steps in machine perception and hence performs very robustly.

The related research is carried out at Jacobs Robotics as part of the EU-project “Cognitive Robot for Automation of Logistics Processes (RobLog)”.