What’s Hot in Robotics?

Artificial Intelligence, Machine Learning, Neocortex, Robotics, Robotics Business / 18.06.2015

Frank Tobe’s article in Robohub regarding what’s hot in robotics reads like a who’s who of companies providing state-of-the-art applications in robotics. He breaks down the trends into four buckets, one of which I’d like to address – Advances in Visual Perception.

Here’s what he says:

“Vision-enhanced robotic systems are becoming the top reason for upgrading and deploying vision-enabled robots and a core reason for the steady upward growth of the robotics industry…. Artificial intelligence and various AI learning systems have been improving regarding visual perception, and many new companies (such as Universal Robotics and their Neocortex system) are now either offering vision systems that can supplement existing fixed systems or offering mobile manipulators that can find and determine how best to pick and handle all sorts of objects from plastic-wrapped toys to boxes, cases and skids of materials.”

Frank’s points are correct. Let’s unpack the reasons for this burgeoning market opportunity.

Traditional vision systems, working within a set of predefined attributes, are superb at locating objects for manipulation. These systems require preloading the object’s physical metrics off-line in order to function and are inherently incapable of change without additional manual software changes.

Flexible 3D vision systems build upon the traditional vision paradigm to include variability in objects and their positioning without manual software changes off-line.

For example, such variability could include a dented box or one that lost its label. Objects with similar visual attributes, like never seen before large pill bottles, could be recognized because they are similar to smaller bottles.

These flexible systems are useful for supply chain activity, which has random variability in object placement, shape, and high SKU count. Case in point: mass-customization is becoming the norm as manufacturers continue to increase the variability in their product offerings. Consumer demand requires retailers to maintain inventories with greater SKU count to service fast delivery.

Vision solutions, whether fixed or flexible, use probabilities to decide if what is being seen is, in fact, the object sought. A variety of algorithms can be combined in the effort to make a judgment. Techniques include extracting and matching discreet features, matching shape based on a CAD input, or matching surface geometry. But what happens when they fail to return a probability that will lead to success? If for example, in a bin picking application, the vision system fails to recognize the next object to be picked, then what?

This is where artificial intelligence takes over. Techniques in machine learning provide methods that improve probabilities by coupling the software’s previous experiences with current observations. There are a host of methods, which when taken together, improve outcomes and direct the vision system to generate the correct choice. For example, if vision cannot distinguish the outside edge of a box from the seam between the flaps, AI will use prior knowledge of box characteristics to inform its judgment.

This is great, but aren’t these systems slow? No. And this is the last piece to make these solutions viable for industry use. The algorithmic gymnastics required for supply chain throughput must exceed human capability to be useful. This is now possible. We are at the intersection of two mega trends: software is becoming more sophisticated and efficient, and computation power continues to increase exponentially. Contrary to Amazon’s recent bin picking challenge, where the winning entrant picked two objects per minute with 83% reliability, Universal is delivering the speed required with high reliability.

Smart robotic 3D vision is very hard to do. Many have attempted it, only to discover a very high barrier to entry. The Universal software platform takes full advantage of the state-of-the-art processing, algorithm design, sensor type, and advanced robot motion control. It combines 3D perception with machine learning to drive robot control, typically in less than 700 milliseconds.

The advantages to software solutions are well understood. Software provides the opportunity to move complexity from the mechanical to the digital. Today, these software-driven solutions reduce fixturing costs, increase the ability to adapt to changes quickly, and support mass customization while driving down costs.

Indeed, as Frank states in his article “Four diverse trends are having a big effect on the robotics industry worldwide, and the global media and financial gurus are paying attention to the process.” Indeed they are, and with good reason.

Tags: , , ,

Comments are closed.