Amazon 原型機以機械人操作減少物件損壞

Originally published on The Daily Flare.

Robot manipulation is moving beyond simple pickup: an Amazon prototype uses three-dimensional scene understanding and pinch grasps to handle selected items with less damage in preliminary tests.

Reading a cluttered scene

Picking up an object is only one part of handling it. A robot also has to find the intended item, work out how it sits among other objects, choose a place to make contact and keep control while moving it. Those steps become difficult when shelves or bins are crowded, or when other objects block part of the target.

Amazon Science describes research aimed at helping robots understand cluttered environments in three dimensions, identify specific items and retrieve them safely. The task resembles finding a bottle in a crowded refrigerator without knocking over nearby containers. A person can often infer an object’s shape from a partial view, using experience and surrounding context. For a robot, building a useful representation of the scene is a central part of the problem.

Amazon Robotics AI director Siddhartha Srinivasa described the work as combining three-dimensional scene models with item identification and machine learning. The goal is not simply to detect that something is present. A robot needs to identify which object it should move and select a way to grasp it that suits its shape and position.

That makes robotic manipulation different from a basic lift-and-place action. The robot must manage an object through movement, not merely make contact and raise it. Amazon applied scientist Can Erdogan emphasized this distinction in describing the research: controlling an item as it travels is essential to moving it reliably.

Robotic grasping is therefore one component of a larger handling process. A grip that initially succeeds can still fail if the object shifts, swings or slips while being carried. The robot’s understanding of the surrounding scene and its ability to plan a controlled movement matter alongside the physical design of its gripper.

Amazon’s researchers frame the broader challenge as item-specific manipulation. Fulfillment centers handle objects with many different shapes and sizes, so a method that works for one package may not work for another. The company’s stated ambition is to develop systems capable of handling millions of different items, but the account describes that as an ongoing research effort, not a completed capability.

Pinch grasps and the limits of suction

Vacuum suction is a common approach for robot pickup because an elastic cup can conform to a surface and create a seal. Amazon Science notes that suction can handle a range of objects, including gift cards and cylindrical poster tubes. For many items, the approach offers a practical way to lift and move them without a complex hand.

But suction depends on keeping the seal intact. If the attachment angle changes during movement, the seal can break. Erdogan described how a quickly moving object may swing away and fall. Increasing suction to reduce the chance of detachment can create a different problem: pressure may blister or tear packaging.

Some objects also require contact with more than one surface. A book lifted from only its front or back cover may flop open, while a bag of loose material can be difficult to seal tightly. These examples show why a single pickup method cannot reliably suit every shape, surface and movement.

A pinch grasp uses two points of contact, in a motion similar to holding an item between a thumb and finger. Rather than relying on a vacuum seal, the gripper holds the object between surfaces and aims to keep it stable as it moves. The method can be useful when suction struggles with the item’s form or when control during transfer matters.

The approach does not mean the robot has a human hand. Aaron Parness, a senior manager for applied science at Amazon Robotics AI, said the field does not have the mechanical ability of a five-finger dexterous hand. He described the research direction as pairing a mechanically simple gripper with improved reasoning about how to grasp. In that sense, robot dexterity can come partly from better decisions about contact and movement, rather than from copying the full anatomy of a hand.

Amazon Science reported that preliminary tests of the prototype showed a 10-fold reduction in damage on items such as books and boxes. The result applies to those early tests and examples. It is not presented as a universal measure across all objects, nor does the account establish that the prototype is already handling the full range of items in an operating fulfillment center.

Precision depends on the whole movement

The reported damage reduction points to a broader goal in robotic manipulation: a successful grasp should preserve the object and maintain control throughout its transfer. That requires a system to connect what it sees with a choice about where to grip and how to move. A strong hold alone is not enough if it damages a package or makes the item unstable.

Amazon’s research combines scene understanding with machine learning to help select a grasp and move an identified item. The source account does not detail the model architecture or provide a full testing protocol, so the precise technical steps behind the prototype’s decisions are not established there. What it does describe is the intended relationship between perception, grasp selection and controlled motion.

That relationship matters in clutter. The target may be partly hidden, and nearby items can constrain how the robot approaches it. A robot needs to distinguish the intended object from its surroundings, then use a grip that works in the available space. The three-dimensional model helps represent the clutter, while machine learning is used to reason about how best to pick up and move the selected item.

Pinch-grasping is one answer to a specific set of problems, not a replacement for suction in every situation. Suction remains useful across many shapes, while pinching can address cases where a seal is unreliable, contact on multiple surfaces is needed or a controlled hold is preferable. A capable system may need to choose among approaches rather than depend on one gripper action for every object.

The central engineering challenge is to make that choice dependable across varied objects without adding unnecessary mechanical complexity. Parness’s description captures the trade-off: a simple mechanism can become more functional when paired with better reasoning. For robot dexterity, the important advance is not only a more elaborate hand, but also the ability to understand a scene and adapt a grasp to the object.

The prototype’s early tests offer a measured example of that direction. Amazon Science reports reduced damage for selected items, while the wider goal of handling a vast variety of goods remains a research challenge. The work illustrates how precise object handling depends on perception, grasping and movement working together, with each part judged by whether the robot can carry the item safely to its destination.