The Amazon Picking Challenge was a competition (running 2015-2017) to automate one of warehousing’s stubbornly hard problems: reliably picking individual items off cluttered shelves and stowing them into bins. Robots faced a wide variety of consumer products in unstructured arrangements - exactly the conditions that defeat traditional industrial automation, which assumes parts arrive in known positions. The 2016 edition added a stowing task alongside picking, raising the difficulty.
Team Delft, a collaboration between the TU Delft Robotics Institute and the company Delft Robotics, won both the picking and stowing competitions in 2016. Their system paired an off-the-shelf industrial robot arm with 3D cameras and a custom gripper, and used deep learning for object recognition and pose estimation, plus dedicated grasp planning and motion planning. This documented win, published as an arXiv paper submitted October 18, 2016, marked the arrival of learned perception (riding the post-ImageNet wave of deep vision) into practical robotic manipulation.
The challenge matters because grasping arbitrary objects is a deceptively hard problem that sits between AI and the physical world: perception, geometry, and control all have to work together on messy, real items. Amazon used a public competition to accelerate it, and the techniques on display fed directly into the warehouse automation that now underpins e-commerce logistics. For a business reader, it is a concrete case of a benchmark contest pulling research toward a problem with immediate commercial stakes.