Bin picking

Bin picking is the task of a robot picking individual objects from a bin or container, often piled at random, overlapping, and in varied orientations. It combines object detection, 6D pose estimation, grasp planning, and collision-free motion, and it is a standard test for industrial physical AI because clutter, occlusion, and shiny or reflective parts make perception unreliable.

What is bin picking?

In bin picking, a robot looks into a container of parts, chooses one it can reach, picks it up, and places it somewhere useful, such as a fixture, a conveyor, or a machine. It is one of the oldest problems in industrial robotics, and it stays hard because the robot cannot count on any part being where the last one was.
Bin picking is often split by how orderly the parts are. In structured picking, parts sit in known positions. In semi-structured picking, they are separated or roughly aligned. In random bin picking, they are piled in any orientation, touching and hiding each other, and the robot has to work out each part's pose from scratch.

Key takeaways

  • Bin picking means picking individual parts from a cluttered, unordered container.
  • It depends on accurate 6D pose estimation, grasp selection, and collision-free motion.
  • Shiny metal parts, clutter, and occlusion are the classic failure cases for the vision system.

How it works

A 3D camera above the bin captures the scene. The vision system segments the visible parts and estimates each one's 6D pose. A planner scores candidate grasps by how reachable and collision-free they are, picks one, and uses motion planning and inverse kinematics to move the end-effector in without hitting the bin walls or other parts. Many systems check the grasp before placing the part, because one bad pick can jam the next step.
Bin picking chains capture, pose estimation, grasp choice, and a check before placing.

Why it matters

Bin picking automates repetitive work at the start of many production steps, such as feeding parts to a press or a CNC machine, so it is often part of machine tending. It is also where perception errors turn directly into downtime.
Shiny metal is the classic hard case. Reflective surfaces bounce light in ways that corrupt 3D measurements, leaving holes or false points in the depth data, so the pose estimate drifts even when a person can see the part clearly. Clutter adds occlusion, symmetric parts add ambiguous orientations, and near-identical parts make it easy to pick the wrong one. In 2025, the BOP benchmark added industrial tracks focused on bin picking, with cluttered scenes of metal parts.
Improving a bin-picking system usually means collecting its failures: grasps that slipped, poses that were off, and bins where nothing was pickable. Those cases are rare in a large log of successful picks, so finding them is a failure mining problem as much as a modeling one.
Reflective parts leave gaps in the depth data, and the pose estimate drifts.

Frequently asked questions

What is random bin picking?

Picking parts that are piled in a container in any orientation, overlapping and partly hidden. It is harder than structured picking because the robot must estimate each part's pose before it can grasp it.

Why are shiny parts hard to pick?

Reflective surfaces bounce the camera's projected light away from the sensor or onto other parts, which leaves gaps or false points in the depth data. The pose estimate then drifts, and the planned grasp misses.

What sensors are used for bin picking?

Usually a 3D camera mounted above the bin, such as a structured light or stereo sensor, sometimes paired with a camera on the robot's wrist for a closer look before the pick.
black and white photo of Jesse Mostipak
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 19, 2026

Building visual or physical AI?

Let's talk.