Listen to this article
Robots that pick items from Amazon’s warehouses must be able to handle millions of different items of varying shapes, sizes, and weights. Currently, the company primarily uses suction cup grippers, which use air and a tight seal to lift items, but Amazon’s robotics team is developing a more flexible gripper to reliably pick up. items that suction cup grippers struggle to grip.
Amazon teaches robots how to understand cluttered environments in three dimensions, locate specific items, and select them using a pinch grip or a thumb and finger grip. The company’s current vacuum grippers use elastic suction cups that form on the surface of an item. This creates a tight seal that allows the robot to grab objects.
Amazon said this method works great for flat items that only require one touch point for picking, such as rulers or maps. It’s less effective, Amazon said, for items that require more than one point of contact to retrieve—for example, a book will open if you pick it up from the front or back cover only.
According to Amazon, suction cup clamps also struggle to seal bags full of grainy items, like marbles, tightly. And even on items that these grippers can grip well, if the attachment angle changes due to the momentum of the robot arm swinging it from place to place, the seal will break too soon and the robot will drop the item.
These cases explain why Amazon is interested in the pinch input method. Although it is natural for humans, it is not easy to develop in a robot. To teach a robot to choose objects from piles of other objects using this method, researchers first had to teach it to be able to estimate the shape of objects that might be partially obscured by other objects. .
As humans, we do this without even thinking about it, but robots have a much harder time understanding the full shape of an object if they can’t see everything. Amazon’s robots evaluate what they choose using multiple camera angles and machine learning models trained to recognize and estimate the shape of individual items. The robot uses this to decide how best to grip the object on two surfaces.
Once the robot has made these observations, it uses a set of motion algorithms to combine the information it has gathered about the scene and the item with the robot’s known dynamics to calculate how to move the item. one place to another.
The robot also continues to use its multi-angle view of the situation throughout the selection. This is another departure from typical selection methods, where a robot usually does not keep looking at the scene while making a selection.
So far, the Amazon team has had encouraging success with their pinch robots. A prototype robot has reduced damage to certain items, such as books, by 10 times without slowing down operations, Amazon said.
Despite this, Amazon still sees room for improvement. The team currently uses an off-the-shelf gripper that can only grab items weighing less than 2 pounds. This makes the clamp capable of handling only half of the items available for purchase on Amazon. In the future, the team plans to design its own gripper for the job.
In the future, Amazon hopes to be able to implement its pinch-to-grip robot alongside its current vacuum robots so that it can decide which robot would be best suited to pick each individual item. The company uses a similar strategy with its autonomous mobile robot (AMR) Proteus.
Amazon unveiled Proteus in June 2022, expanding its already extensive robotics ecosystem. The company had already been using automated guided vehicles (AGVs) in its warehouses for a decade, but these AGVs only work in caged spaces.
While the two robots perform similar tasks, sliding under Amazon GoCarts to pick them up and move them around the warehouse, Proteus gives the company greater flexibility due to its ability to work freely and safely. with people. The company plans to continue using both robots in the future.
Amazon recently announced that it has agreed to acquire Cloostermans, a Belgian company specializing in mechatronics. Cloostermans has been selling products to Amazon since at least 2019, including the technology Amazon uses in its operation to move and stack heavy pallets and bins and robots to package products for customer orders.