Faits confirmés
Boston Dynamics’ Stretch performs a defined warehouse task: moving boxes from a trailer to a conveyor while deciding in real time how to see, grasp and route around the next obstacle. That boundary matters. The robot is handling useful work in customer sites, but its performance still depends on how reliably its sensors interpret the environment.
Déclarations de la source
Stretch builds a view of the trailer
Analyse de la rédaction
Boston Dynamics presents Ethan Lauer as a perception engineer on Stretch’s world-modeling team. Stretch combines cameras and sensors to interpret its surroundings, handle boxes and navigate around obstacles. Its perception mast uses two cameras that produce standard RGB images and depth point clouds. Four lidars mounted on the base provide a 360-degree planar view for detecting boxes and walls, then navigating toward the next grasp. Several distance sensors in the gripper refine the robot’s view of the ceiling, nearby obstacles and the grasping area. The arrangement gives Stretch different levels of information for different parts of the job: a broad view helps it move through the trailer, while closer readings help position the gripper on a particular box. Every pick involves a decision During a pick, Stretch photographs the scene and selects a box without disturbing the others. When possible, it favors the top of a stack to reduce the risk of the pile falling. The box’s weight and dimensions influence which side the robot grips. Before moving the load to the conveyor, Stretch checks its surroundings for collisions and plans the arm’s path. These steps happen in real time, so the workflow is more than a fixed sequence of identical reaches. Stretch can respond to the box in front of it rather than relying on every item being described in advance. The demonstrated job remains specific, however: unloading boxes from a trailer and sending them to a conveyor. It is evidence of a bounded capability, not a claim that Stretch can perform general-purpose work in every warehouse environment. What “plug-and-play” means here At a site, general parameters tell Stretch where the dock is. The remaining operation is described as nearly plug-and-play, with the perception system handling most of the work. On a first encounter, Stretch does not know a box’s type, size or weight, yet it can grasp the box automatically and learn from that grasp. The practical benefit is reduced preparation for each individual box. The trade-off is that the robot still has to interpret a changing physical scene correctly at the moment of handling. Adaptation helps with variation in the load, but it does not remove the need for reliable sensing around the robot. A false obstacle can interrupt the flow In one test, a thin spider web resting on a base lidar was interpreted as an obstacle that did not exist. Stretch moved away from it. The incident gives the plug-and-play claim a concrete limit: the system can react conservatively to a misleading sensor reading, even when the perceived obstacle is harmless. That response protects the robot from treating uncertain information as a clear path, but it can also interrupt the movement of boxes. In a warehouse workflow, the sensing error is therefore part of the operational picture. Autonomy is not only about selecting and grasping the next box; it is also about deciding when the surrounding data is trustworthy enough to continue. A deployed workflow with defined boundaries Lauer says he speaks with clients to check that the robots are working and sees them move their boxes. Those conversations and observations feed improvements to the robot. They describe a perception-driven box-handling workflow operating at customer sites, rather than a future plan for a system that has not yet been used. Stretch’s significance is concrete: it can sense a pile, select a box, adapt its grasp and move the load through a trailer-to-conveyor task with limited site-specific setup. The spider-web false alarm is equally useful for understanding the system. Real warehouse autonomy can absorb variation in the boxes, but its value remains tied to the boundaries supported by its cameras, lidars and distance sensors. Official sources Official source: bostondynamics.com Related reading Kawasaki Astorino Industrial Robot Training Classrooms Amazon Millionth Robot Deepfleet Fleet Software Safety
