Robots in Amazon Warehouses: Constant Progress, But Humans Remain Indispensable | Festina Lente - Your leading source of AI news | Turtles AI

Amazon recently conducted extensive tests on two robots designed to streamline operations in its fulfillment centers: Stow, which is dedicated to putting away items, and Pick, which is responsible for picking items. While both robots showed significant progress, the results suggest that full replacement of human labor is not yet on the horizon.

Key Points:

• The Stow robot achieved an 85 percent success rate across more than 500,000 putaway operations, but caused damage to items in 9 percent of cases.
• The Pick robot achieved a 91 percent success rate across 12,000 pick attempts, but rejected 19.4 percent of requests due to difficulty recognizing items.
• The Stow robot’s putaway speed is comparable to that of human workers, at 224 units per hour compared to 243 for humans.
• Amazon has introduced Vulcan, a robot equipped with advanced tactile sensors that can handle about 75 percent of the items in its warehouses, improving its ability to handle irregularly shaped objects.

In Amazon’s fulfillment centers, products are stored in fabric containers that resemble suspended shelves, known as “pods.” When an item arrives at the warehouse, workers assess its quality and manually stow it in the pods. Then, when a customer places an order, the entire pod is transported to a picking station, where the item is selected and sent for packing and shipping.

The Stow robot was designed to automate the putaway process. Equipped with a gripper and an extendable table, it uses a visual perception system to assess available space in the pods and a machine learning model to predict packing success. In tests, it handled more than 500,000 items with an 85 percent success rate. However, 9% of the failures involved damaged items, often due to drops, and 14% saw book pages ruined while being placed in pods. Its putaway speed was similar to that of human workers, with 224 units per hour compared to 243 for humans.

The Pick robot was tested for picking items. During a six-month trial, it achieved a 91% success rate out of 12,000 attempts. However, it rejected 19.4% of pick requests due to difficulty recognizing the items or to avoid possible damage.

Amazon also introduced Vulcan, an advanced robot equipped with tactile sensors that allow it to determine how much pressure to apply when handling items. This innovation allows Vulcan to handle a wider variety of irregularly shaped items, allowing it to pick and put approximately 75% of the items typically found in a fulfillment center. Vulcan is currently operating at sites in Spokane, Washington, and Hamburg, Germany, with plans for broader deployment across the U.S. and Europe in the coming years.

Despite this progress, Amazon stresses that full automation is neither feasible nor desirable at this time. The company emphasizes the importance of humans and robots collaborating to maximize efficiency and productivity in warehouses. Additionally, Amazon is exploring teaching robots through visuo-motor policy (VMP) learning, rather than manually programming specific behaviors. A key challenge in implementing learned VMPs is their poor interpretability when a fault occurs, requiring retraining or tuning to learn from failures while maintaining past performance.

To address these challenges, Amazon plans to improve the performance of VMPs by modeling failures in a Real2Sim module, which involves generating digital replicas of real-world scenes through robotic interactions. This approach should help address rare failure cases that arise at scale.

While robots like Stow, Pick, and Vulcan represent significant advances in warehouse robotics, their full integration into Amazon’s operations requires further development and close collaboration with human workers to ensure efficiency, safety, and reliability.