Boston Dynamics has introduced a new perception system for Atlas, its humanoid robot, to help it work in real-world factories. With this enhancement, Atlas can recognize and understand its environment as precisely as a human, which is essential for performing tasks on its own in factories and other complex areas.
While Atlas was known for being agile and mobile, the focus is now on perception, which helps it navigate busy and changing environments. The robot is now able to identify shiny, dark, and tightly packed objects by using a system that combines 2D and 3D vision. First, a 2D detector finds objects, then it adds bounding boxes and keypoints, and finally, it improves its understanding by tracking poses in 3D.
In factories, Atlas often works with fixtures that hold the parts. Both the outer and inner keypoints are used to analyze these fixtures and find their general shapes and internal slots. Atlas can quickly and accurately estimate where objects are, even if the view is poor or the angles are confusing.
A localization module helps the robot by matching its visual data with stored models and motion input to make its predictions more accurate as time goes on. The system relies on spatial memory to tell apart fixtures that are almost the same.
After Atlas holds a part, it must be tracked in space by the SuperTracker system. It uses kinematic, visual, and force information to keep track of object movement, notice slips, and hold on to the object.
The robot’s vision system is taught using computer-generated data and checked for accuracy with self-consistency and kinematic rules, so what Atlas sees matches what it feels. The robot’s limbs and cameras are calibrated so precisely that it can move with sub-millimeter accuracy.
This new perception system is a big step forward for real-world self-driving cars. Boston Dynamics believes that the future involves creating a single model that combines perception and action into a single ability.