Amazon Picking Challenge aimed at improving warehouse robotics
One of the biggest events at the recent 2015 IEEE International Conference on Robotics and Automation (ICRA) in Seattle was the first Amazon Picking Challenge, in which 31 teams from around the world competed for US$26,000 in prizes. The challenge set entrants with the real-world task of building a robot that can do the same job as an Amazon stock picker.
Internet commerce may have revolutionized shopping, but it still relies on armies of people toiling away in warehouses to get the goods out to consumers, with Amazon alone hiring 80,000 temporary warehouse workers over last year's holiday rush. It's an obvious target for automation, so in 212 Amazon spent US$775 million to acquire robotics company Kiva Systems and installed 15,000 warehouse robots.
According to Amazon, the Kiva robotic system sped things up by reversing the job of stock picking. Instead of the worker going to the shelves and picking out items, the shelves came to the workers, who picked and packed the goods. The shelves even helped them by identifying items and showing workers where they go by means of lasers, and even confirmed the item's bar code before packing.
The new system is faster and more efficient, but having humans pick the items off the mobile shelves by hand is still an expensive, tedious task that's an ecommerce bottleneck. This is what inspired the Amazon Picking Challenge, which seeks to find ways to replace the human stock picker with robots.
According to Amazon Chief Technology Officer Peter Wurman, who initiated the challenge, the task of picking items off the shelf may seem simple, but it involves all domains of robotics. The robot has to capable of object and pose recognition. It must be able to plan its grasps, adjust manipulations, plan how to move, and be able to execute tasks while noticing and correcting any errors.
This might suggest that the robots would need to be of a new, specialized design, but for the Picking Challenge, Amazon made no such requirement. According to one participant we talked to, the more important factors were sensors and computer modelling, so ICRA 2015 saw all sorts of robots competing, such as the general purpose Baxter and PR2, industrial arms of various sizes, and even special-built frames that move up, down, left or right to position the arm. Even the manipulators used by the various teams ranged from hooks, to hand-like graspers, and vacuum pickups.
The rules of the Picking Challenge were based on a simplified version of the warehouse goods picking task. Each team was provided with a stationary Kiva shelf pod consisting of 12 cubbyholes holding 24 items of various sizes and weights, such as books, cat toys, and cookies. Each team was given a list of 12 items and 20 minutes for the robot to select them and place them in a tote bin correctly without damaging them. Points were scored based on how many items were correctly picked against incorrectly picked, dropped, or broken items.
According to Wurman, the challenge went off as planned, though the participants did run into unexpected problems in the transition from computer models and laboratory tests to a more realistic setting. Other observers told us that the robots were much faster than originally estimated, and that the main problem was with the manipulators having to operate in the confines of the cubby holes and at unanticipated angles.
The RBO team from the Technical University of Berlin (TU Berlin) took out the $20,000 first prize, while Team MIT claimed the $5,000 second prize and Team Grizzly from Dataspeed Inc and Oakland University took out the $1,000 third prize.
Please keep comments to less than 150 words. No abusive material or spam will be published.
If anything, this should be an X-prize level challenge because you are solving two of the most challenging issues in automation today.
The technology just isn't there yet for pickers to be a direct replacement for humans so I'm sure Amazon set the prize low when they set the bar low.
One solution I can think of is to identify a small handful of items the robots would be able to pick reliably and store those ~20-30 items in a separate section of the warehouse. Depending how commonly sold those items are it could lessen the burden on human employees having robots handle just those. As technology improves maybe that becomes 50-60 items instead.
Telling some of the things apart and stocking them in a way a robotic picker could select them (Fire TV stick vs 50 shades of grey etc.) would be fairly trivial. It may be a case where Kiva brings over a shelf full of only copies of a single item leaving less room for error of selecting the wrong item. What percentage of their pickers time is spent picking 100-200 of their best selling items? It's probably significant.
I suspect there is probably some amount of labor (like 30%) where it becomes worth the cost to deploy a system like this and start improving on it in iterations.
This 80 000 temp jobs which will eventually be lost is just a drop in the bucket.....