Artificial intelligence is enabling robots from KNAPP to fulfil orders quickly and accurately at a European distribution centre.
Equipped with leading-edge AI, the Pick-it-Easy Robot solution is picking customer orders in a real-world application day in, day out. What’s more, it is doing this with 99% accuracy and at a speed equivalent to a manual picking station – although, of course, it can maintain this speed over a much longer period than a human picker.
Robots for retail
“Advances in AI are enabling robots to progress from repetitive to random tasks,” says Phil Houghton, Business Development Director for KNAPP UK. “This means that robots can learn quickly how to grasp unfamiliar objects, which has long since been the Holy Grail in retail logistics.”
While humans can grasp a huge range of items without even thinking about it, robots have to work out how to handle an unfamiliar object. There are many characteristics that can confuse them – such as size, shape, weight, surface type, packaging or the physical position of items. This presents a challenge when it comes to fulfilment, where there is an ever-changing range of SKUs and items are arranged chaotically in totes. As a result, until now robots in the warehouse have tended to be confined to repetitive, uniform tasks such as depalletizing and palletizing. Now, advances in AI mean that robots can learn much more quickly and dynamically. “What’s more,” says Phil Houghton, “as the number of robots in distribution centres continues to rise, this learning increases exponentially because robots around the world are sharing their learning with each other.”
Deep learning
Just like people, robots learn by experience – every time a robot tries to pick up an item, what it learns is generated in the form of data and can be applied to other articles. KNAPP has collaborated with Californian start-up, Covariant, to develop its Pick-it-Easy Robot solution. The US company has pioneered what it terms the ‘Covariant Brain’, based on various types of AI: reinforcement learning, imitation learning and meta-learning. This enables the Pick-it-Easy Robot to learn how to handle unfamiliar items by excelling in tasks such as 3D perception, few-shot learning and real-time motion planning.
What is involved in these techniques? In reinforcement learning, a robot trains itself through trial and error. This learning can also be performed in simulation, to speed up the process. Imitation learning involves robots learning to copy a human in completing a task – not by exactly repeating movements, but rather by making generalisations about what is observed. Reinforcement learning and imitation learning can be combined in a hybrid system. Here, a human instructor uses a virtual reality headset and controllers to teach a robot to perform a particular task, and then reinforcement learning enables the robot to refine its performance via trial and error. Meta-learning could be described as ‘learning to learn’; it involves refining the learning algorithm itself through experience gained from multiple learning events. “Rather than learning how to master specific tasks,” says Phil Houghton, “the Pick-it-Easy Robot learns general skills that it can apply to new tasks, in much the same way that people do.”
No better time to invest
The Covid-19 pandemic has boosted demand for e-commerce. With e-com orders requiring more processing in the warehouse compared with orders for store replenishment, there has been a sharp uptick in interest in logistics automation, robotics and AI to boost productivity and accuracy. With the cost of robotic technologies having fallen considerably in recent years and with low interest rates currently available, now is an attractive time to invest in robotic automation. “What’s more,” adds Phil Houghton, “we offer our Pick-it-Easy Robot in a Robotics-as-a-Service (RaaS) model, giving clients the efficiency benefits immediately but spreading the investment cost over time.”
Comments are closed.