An MIT student improves robots' abilities to recognise objects and their orientation in a project designed to teach them table tennis.
The algorithm, designed by Jared Glover, a graduate student in MIT’s Department of Electrical Engineering and Computer Science, exploits a statistical construct known as the Bingham distribution. Glover has created an open-source software tool designed to speed up the Bingham distribution calculation while analysing orientation of ping-pong balls in flight, as a part of a broader project to teach robots to play table tennis.
The algorithm offers a particular improvement in robots' visual recognition performance in settings where visual information is particularly poor, performing 50 per cent better than its competitors.
Compared with its rivals, the system is 15 per cent more effective at identifying familiar objects in cluttered scenes. However, the creators believe the nature of the Bingham distribution will enable making it even more effective in settings where information is patchy or unreliable.
“Alignment is a key to many problems in robotics, from object-detection and tracking to mapping,” Glover said. “And ambiguity is really the central challenge to getting good alignments in highly cluttered scenes, like inside a refrigerator or in a drawer. That’s why the Bingham distribution seems to be a useful tool, because it allows the algorithm to get more information out of each ambiguous, local feature.”
The major strength of the algorithm is its ability to combine information from various sources. Generally, determining an object’s orientation requires superimposing a geometric model of the object over visual data captured by a camera – in the case of Glover’s work a Microsoft Kinect camera, which captures a 2D colour image together with information about the distance of the colour patches.
The software can identify locations in the image where colour or depth values change abruptly – likely to be the edges of an object.
The problem is then reduced to taking two sets of points – the model and the object – and determining whether one can be superimposed on the other. Most algorithms, Glover’s included, will take a first stab at aligning the points.
If both sets of points in fact describe the same object, then they can be aligned by rotating one of them around the right axis. For any given pair of points – one from the model and one from the object – it’s possible to calculate the probability that rotating one point by a particular angle around a particular axis will align it with the other. The problem is that the same rotation might move other pairs of points farther away from each other.
Glover was able to show that the rotation probabilities for any given pair of points can be described as a Bingham distribution, which means that they can be combined into a single, cumulative Bingham distribution.
The Bingham distribution can also incorporate probabilities from other sources of information, such as estimates of the curvature of the objects’ surfaces.
In test, robots using the new algorithm were able to find 73 per cent of objects in a cluttered scene, while their competitors scored only 64 per cent. Glover believes the difference is due to his algorithm’s better ability to determine object orientations.
Incorporating additional sources of information, such as the probability of certain objects to be found in certain positions, could further improve the algorithm’s performance, Glover believes.
“You can spend your whole PhD programming a robot to find tables and chairs and cups and things like that, but there aren’t really a lot of general-purpose tools,” Glover said. “With bigger problems, like estimating relationships between objects and their attributes and dealing with things that are somewhat ambiguous, we’re really not anywhere near where we need to be. And until we can do that, I really think that robots are going to be very limited.”
The results of the research will be officially introduced at the International Conference on Intelligent Robots and Systems in November this year.