Amazon is actively beta testing robots in its fulfillment warehouses to handle tasks like picking and stowing products. While these robots show promise, they aren’t quite ready to operate without human oversight. The e-commerce giant is cautiously optimistic, but human workers remain essential to the process.

Inside Amazon’s Fulfillment Centers

Amazon’s warehouses utilize a unique storage system. Products are housed inside fabric pods, resembling IKEA’s SKUBB storage units, suspended on racks.

The fulfillment process involves several steps:

• Receiving: Human workers assess the quality of incoming items.
• stowing: Using conventional material handling equipment, workers place items into the storage pods.
• Picking: When an order is placed, workers move the entire pod containing the desired product to a picking station.
• Packaging and Shipment: Staff remove the item from the pod and prepare it for shipment.

“Stow” and “Pick”: Amazon’s Robotic Recruits

amazon is currently testing robots named “Stow” and “Pick” to automate these processes. Early results, detailed in research papers, offer a glimpse into the capabilities and limitations of these machines.

The Stow Robot: Packing with Precision (Mostly)

The Stow robot is equipped with:

• A pinch gripper and extendable plank for item manipulation.
• A visual perception system to assess available space in bins.
• A machine learning model to predict packing success.
• Adaptive capabilities to create space for new items.

During a test involving over 500,000 items, the stow robot achieved an 85% success rate.

The Book-Breaking problem

However, the Stow robot isn’t perfect. Failures included:

• 9% resulting in item damage,often from items falling.
• 14% involving mangled book pages.

The latter issue has prompted Amazon’s roboticists to consider “book-saving interventions.”

Speed comparison: Robot vs. Human

The Stow robot’s speed is comparable to that of human workers. Over the month of March 2025, humans stowed at an average rate of 243 units per hour (UPH) while the robotic systems stowed at 224 UPH, according to the research paper. The comparison focused on human stowers operating on the same floor as the robotic workcells to ensure a fair assessment, as stow rates vary based on inbound item distributions and the density of items already in the fabric pods.

Researchers believe human stow rates could increase by 4.5% if robots handled the top rows of storage pods, eliminating the need for ladders.

The Pick Robot: A More refined Approach

The Pick robot demonstrated a higher success rate during its six-month trial, operating six hours per day on weekends. It achieved a 91% success rate across 12,000 pick attempts.

however, it also rejected 19.4% of pick requests due to:

• Machine vision failures in recognizing the item.
• Concerns about potential item damage.

The Future of Robotic Learning at Amazon

Amazon’s researchers view these experiments as crucial steps toward deploying reliable robot systems at scale. The company intends to explore visuomotor policy learning (VMP) to teach robots,rather than relying on manual programming.

However,VMP presents challenges.A key challenge of deploying learned VMPs is their lack of interpretability in failure cases, Amazon’s researchers observe.While it is often easy to fix a simple bug or improve algorithmic limitations of heuristic approaches, VMPs have to be re-trained or fine-tuned to learn about failures while maintaining their previous performance.

To address this, Amazon plans to improve VMP performance by modeling failures in a Real2Sim module, creating digital replicas of real-world scenes through robotic interactions. This approach aims to resolve rare failure cases that emerge at scale.