Robots Taught to Think Before Acting

Thanks to a new computer algorithm robots are now able to move in and around cluttered environments.
Fabienne Lang
The theoretical work has focused on using quantum computing to accelerate the machine learning/SINCSource: Eurekalert

Robots are increasing in their numbers in our daily lives, by taking over simple tasks around homes and in businesses. During their tasks, these robots are faced with a wide span of articulated objects, such as tools, cabinets, drawers, and other jointed objects. These objects offer an infinite number of possible arrangements and pose, and robots have to quickly discern all possible variations in poses to move or retrieve objects in these spaces.


The problem remains as to how to teach a robot all of these possible variations of poses, and how it can navigate its way through cluttered, unplanned environments. 

In order to improve robotic perception, scientists from the University of Michigan at Ann Arber, led by Karthik Desingh, have created a methodical algorithm which calculates and estimates various potential poses of jointed objects. Condensed, the algorithm teaches the robot to coordinate its actions. The article about the research "Efficient Nonparametric Belief Propagation for Pose Estimation and Manipulation of Articulated Objects" is published in this week's Science Robotics journal.

Robots Taught to Think Before Acting
Envisioning the future of robotics. Source: Ryan Etter/Robohub

Robots thinking before acting

Ultimately, this will be extremely useful and will improve service of warehouse robots or home robots, as their ability to interact and move around jointed objects and tools will drastically increase. 

In order to move quickly around a kitchen and its cabinets, a robot has to comprehend and know the cabinet's range of poses (closing and opening drawers, for example), by practicing a specific set of movements. 

The challenge? Variation in the norm.

For example: If a kitchen towel is strewn across drawers or a cabinet, the robot no longer recognizes the object and does not know which move to make next.

Thanks to the new algorithm a robot will now be able to take this into account, run through all possible pose variations and still be able to go around it and figure out how to work in and around the cluttered environment. This was not the case previously. 

Robots Taught to Think Before Acting
A Fetch robot faces a cabinet, estimating the pose of the cabinet in order to operate on it. Source: Desingh et al., Sci. Robot. 4, eaaw4523 (2019)

Understanding the algorithm

Desingh and his colleagues created the algorithm, named PMPNBP, which formulates random variables that constitute different options of a sequence of pose assessments. It uses the robot's prior understandings to do so.

At present, 100 distinct iterations are used through PMPNBP, leaving space for many a dishtowel to be flung in a cabinet's way.

Robots Taught to Think Before Acting
(A-E) Observed scene, point cloud observation, beliefs of various possible pose estimations at iteration 0, and beliefs at iteration 100 of an open cabinet. (G-K) Observed scene, point cloud observation, beliefs of various possible pose estimations at iteration 0, and beliefs at iteration 100 of a cabinet with a blanket on it. Source: Desingh et al., Sci. Robot. 4, eaaw4523 (2019)

The key to PMPNBP's success? Its researchers have stated that it is due to its partial observations to revolve the entire object of hypothesized posses. It is more precise and systematic when poses of jointed objects are being estimated, jumping a notch ahead of PAMPAS, a pre-existing method.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board