Automating Returns

Returns handling has long been one of retail's most labor-intensive workflows, and one of the least optimized. Each item requires individual handling, physical inspection, condition grading, and a routing decision. At scale, the per-item cost of that process frequently approaches or outright exceeds the item's remaining resale value.

This is why many retailers have historically treated returns as a write-off. Not because they lacked the capability to assess returned goods, but because the economics didn't justify the effort. Discarding items was often cheaper than processing them.

That approach is no longer viable. Return volumes have grown to the point where the inbound flow alone creates significant bottlenecks in warehouse operations. Across Europe and the U.S., logistics providers are now dedicating entire facilities exclusively to returns handling.

What makes returns particularly resistant to automation is the nature of the task itself. Unlike conventional pick-and-place operations, which involve moving a known item from a defined location to another, returns handling is a multi-step decision workflow operating under continuous uncertainty. Every item is different and their conditions vary. Packaging is damaged, missing, or misleading. The correct action at each step depends on what was learned in the steps before it.

No single subtask is the bottleneck. The challenge is managing the full sequence reliably, at speed, across an unpredictable stream of items.

Inside the Returns Workflow

From the moment a customer ships an item back to the point where it's restocked, refurbished, or discarded, the item passes through a sequence of tightly coupled steps. Each one introduces its own complexity, and each one depends on the decisions made before it.

At every stage, the robot isn't just executing; it's deciding. Reshelve, refurbish, or discard? Is this item actually unused, or just cleverly repackaged? Should it attempt a different viewing angle before committing to a grade?

These are sequential decisions under uncertainty, and getting them right has a direct impact on the bottom line.

Built for Contact-Rich Manipulation

Returns handling demands a type of manipulation that single-arm systems simply can't deliver. Opening packages, unfolding garments, and inspecting items all require both arms working in coordination. These tasks can't be broken down into sequential single-arm steps without either adding fixture support or fundamentally redesigning the workflow.

This is what separates returns handling from conventional pick-and-place automation, where a single end-effector operates on predictable, geometrically constrained objects. Returns introduce deformable materials, unknown object states, and contact-rich manipulation sequences that need continuous force distributed across two arms.

Manipulation skills developed in simpler settings carry over as reusable building blocks in more complex bimanual tasks: grasp stability, contact-aware motion planning, dynamic object tracking all transfer. As the complexity of tasks grows, learned capabilities stack rather than requiring the system to start over.

Visual Intelligence for Condition Classification

The Lens system is our visual intelligence layer, suitable for the inspection and classification demands of returns handling. Where Cortex controls what the robot does, Lens determines what the robot knows about each item: its condition, its category, and what should happen to it next.

Manipulation skills developed in simpler settings carry over as reusable building blocks in more complex bimanual tasks: grasp stability, contact-aware motion planning, dynamic object tracking all transfer. As the complexity of tasks grows, learned capabilities stack rather than requiring the system to start over.

Critically, Lens doesn't operate in isolation. It works directly with Cortex's manipulation policies. If a visual inspection is inconclusive because a label is partially occluded or an item is folded in a way that hides a defect, Cortex can actively reposition the item to give Lens a better view. The robot doesn't just look at what's in front of it; it manipulates the scene to gather the information it needs.

Capture

Multi-angle Views

Lens Analysis

Condition Inference

✓ Unused⚠ Dirty✗ Damaged

Classification

Grade & Confidence

Routing

Reshelve / Refurb / Discard

1 / 5

This tight coupling between perception and action is what makes the system work. Lens turns what the robot sees into a decision, and Cortex turns that decision into physical execution.

Why Cortex 2.0 Is Built for This

Returns handling is the kind of problem Cortex 2.0 is well-suited for: long-horizon, high-uncertainty, and deeply tied to real-world variation that no controlled environment can fully replicate.

The result is a system that improves with every shift, learning from the full complexity of real-world returns rather than simplified lab setups or hand-crafted rules.

What Cortex Unlocks Next

Returns handling represents a class of manipulation where success depends on reasoning across extended sequences. Objects arrive in unpredictable conditions, requiring inspection, physical interaction, and conditional decision-making as execution progresses.

Solving it well establishes a foundation for more complex domains like kitting, assembly, and manufacturing, where environmental diversity and physical constraints intensify but the need for sustained reasoning remains the same.

Returns handling marks a critical transition: from robots that excel at single-step tasks to systems that understand and manage processes. Cortex 2.0 is designed for exactly that shift.