How Stretch unloads a trailer one decision at a time

Inside a freight trailer, the next box is not necessarily where a programmer expected it to be. Removing one changes the remaining stack; another may slide or fall. Boston Dynamics’ Stretch robot is designed to keep assessing that changing scene, choosing grips and moving boxes onto a conveyor. The operator assigns an unloading job. The machine works out the individual actions.[2]

There is a commercial history behind that description. DHL introduced Stretch into North American operations in 2023. In its 13 May 2025 announcement of an expanded partnership, DHL reported unloading rates reaching 700 cases an hour and deployments extending into the UK and Europe. The performance figure is the customer’s reported upper rate, without the operating detail needed to treat it as an average across every trailer.[1]

Several boxes, individually controlled

Stretch’s multipick software evaluates candidate groups of boxes, considering their size and how well their faces line up against its suction gripper. Once it chooses a group, the controller tracks which suction cups belong to each box. It checks the seal for each one, then chooses how to place the group. The same cup-to-box mapping allows individual boxes to be released separately.[3]

That last action connects the pick to the rest of the warehouse. Boxes need useful spacing and orientation for scanning and conveyance. Stretch can lift several together, then release them in a way that preserves an orderly flow. Large boxes may require the whole gripper, so the system continues to mix single and multiple picks within the same job. Multipick is part of its normal decision-making, rather than a mode the operator selects for each stack.[3]

An unloading job includes getting out again

Boston Dynamics describes another useful behaviour at the end of a container: Stretch reverses out while retracting the telescopic conveyor and collecting missed boxes along the way. An operator then moves it to the next dock and starts another job. Automating that final withdrawal matters because otherwise the worker has to walk into the emptied container to retrieve the equipment.[4]

A company case study at clothing retailer Gap describes the human handover. Staff drive the robot into position with a wired controller, then use a tablet for the safety check and start. The robot uses machine learning to recognise boxes and the surrounding walls and ceiling. A separate lidar-based perimeter stops motion if someone enters the operating area. This is a specific division of work: people position and start the system, perception and handling software run the unload, and the protective system governs access around moving machinery.[5]

The difficult part can be the cardboard

The manufacturer’s account of commercialising Stretch describes learning from damaged floors, humidity, dust and packaging in worse condition than expected. Its tests include damp cardboard, crushed boxes and weak tape, alongside different conveyors, case sizes and weights. Those details matter to an unloading machine: successfully recognising a box does not guarantee that its surface will support a dependable grip. The product has to cope with the physical state of the thing it has recognised.[6]

The test laboratory also recreates customers’ dock layouts and uses freight matched to their specifications, sometimes the actual products shipped in for testing. Different container clearances test whether the arm can adapt its motion to the available space. Separate reliability cells repeat handling movements for extended periods to expose equipment failures. The evaluation therefore reaches beyond whether one favourable stack can be emptied successfully.[4]

DHL’s 2025 statement described integrating unloading with conveyors and palletisers, and identified case picking as a further development priority. The cited operational and engineering accounts come from the customer and supplier, without an independent comparative trial.[1]

For understanding AI autonomy, the striking feature is the scale of the instruction. Nobody needs to prescribe the next hundred grips. The useful unit of delegation is an unloading job, with the next physical decision generated from the scene in front of the robot. That is how a narrow capability becomes sustained work.[2]

Report an error or suggest a correction