How warehouse roles carried out by large numbers of Amazon's 220,000 staff are expected to be automated in the coming years.
While Amazon employs more than 220,00 people and is growing fast, it is also pursuing technologies that would ultimately allow it to replace many workers in its warehouses worldwide.
Amazon's warehouses - known as fulfilment centers - are already partially automated, with a growing number of centers using Kiva robots to carry shelves of products to human workers, who then pick the items to be shipped.
Without automation, Amazon would be unable to ship items to millions of people each day and as the retail giant moves towards its goal of delivering packages within 30 minutes it needs to continue to streamline delivery.
Not only does Amazon plan to use Kiva bots in more of its warehouses but it also wants to automate the process of picking items from shelves. To that end, the firm last year set up the Amazon Picking Challenge (APC), a contest where robotics researchers compete for a $25,000 prize for designing the best picking bot.
The inaugural contest highlighted just how difficult bots find such tasks, with about half of the teams competing failing to score a single point and the bots working far more slowly than humans. A human can sort about 400 items per hour with minimal errors, while the best robot in the APC achieved a rate of about 30 sorts per hour with a 16 percent failure rate. This relatively poor performance occurred despite the robots being given a far simpler challenge than they would face inside an Amazon warehouse, with bots being tasked with picking from a pool of just 25 items, rather than millions, and with only a few items mixed together.
However, looking back at the contest the teams that took part in the first APC are upbeat - forecasting these early issues will be overcome and that warehouse automation is on the cusp of increasing significantly.
"Recent developments in robotics have the potential of substantially increasing the degree of automation in warehouse logistics and order fulfillment in the near future," according to the paper written by key APC team members - including Peter Wurman, the technical co-founder of Amazon-owned Kiva Systems - now Amazon Robotics.
"The kind of warehouse logistics addressed in the APC, however, can believably be automated using existing or near-future technologies and potentially faster than many other target applications of robotics."
But the teams also warn of the need for "substantial scientific progress" before picking can be automated - pointing out that standard approaches to building robotic hands and creating motion planning software are not sufficient to succeed in the APC.
Building a better warehouse worker
Robotics researchers across the world are working on automating warehouse picking, with advances in many areas. Last week, a team of European researchers revealed a system that picks items within 24 seconds - another step closer to the 5 - 10 seconds it takes human workers to pick each item in an Amazon warehouse.
In contrast, in the Amazon Picking Challenge each robot had more than one minute and a half to pick each item. However, the European bot is working in a very different, and arguably simpler, environment than an Amazon warehouse, with the robot transferring cans between pallets, rather than many different-shaped items from packed shelves.
Nevertheless the European gripping arm - which forms part of an autonomous forklift named the Autonomous Picking & Palletizing (APPLE) platform - demonstrates a novel approach to grasping items, one of the difficulties that contestants had in the APC.
APPLE grips items using the tips of its arm's two "fingers", which close around the item until sensors in the fingertips tell the bot not to exert any more pressure. Conveyor belts on the fingers then roll the object back into the grasp of the hand. The system, which can work safely alongside humans, gets around the difficulty of precisely calculating to what extent the hand should be closed in order to get a firm grip on an object- overcoming another common difficulty with picking bots.
Robert Krug, of the Sweden-based Centre of Applied Autonomous Sensor Systems, was one of APPLE's programmers, and sees their work as progress toward widespread warehouse automation in the near future.
"Clearly there is a discrepancy with a human's speed but there is a big step up from what we could previously achieve with this kind of system."
Krug anticipates picking systems of the type Amazon is pursuing being introduced to warehouses around the start the next decade.
"I don't know whether the first real commercial system will hit in three or five years but it will happen in the very foreseeable future, I'm sure about that."
"The strong interest [in greater warehouse automation] by companies like Amazon shows you the big players think there is something that will happen soon, otherwise they wouldn't put this effort and money into this."
The introduction of automated pickers and other new computerised systems to warehouses will take place gradually - Krug believes - with smaller groups of human workers initially being kept on site to keep the machines on track until they can consistently match or surpass human performance.
"The robot might be able to fulfil its tasks 98 percent of the time, but what do you do about the remaining two percent? One way out is to still have a reduced numbers of humans in the loop. Maybe you could have emergency helpers, who fix tasks when the robot gets stuck."
When asked whether a greater reliance on robots will lead to job losses, Amazon has pointed out that so far its workforce has continued to grow rapidly as it has increased automation.
However, Amazon hasn't explained the commercial benefit of retaining the same number of employees while also running an automated picking system. When asked to comment on the issue for this article, Amazon did not reply.