How reliably do robots pick from a bin?

Not perfectly. A state-of-the-art suction system succeeded on 82.3% of real-world picks, about one in six failing, and only 73.3% on hard objects. Measured rates run below flawless demos.

How fast can a bin-picking robot go?

A low-cost system reached up to 600 picks an hour, best case about 616, at 96 to 99% grasp success on densely packed metal parts. Speed and success depend heavily on the objects.

Why do the two studies give such different numbers?

They measure different things, success rate versus throughput, on different object sets. A high picks-per-hour figure on easy parts is not comparable to a success rate on hard ones.

Do these rates apply to my parts?

Only as a guide. Bin-picking success and speed swing with object difficulty, so easy, rigid, separated parts perform far better than irregular, deformable, or tangled ones.

Robot performance

How well do robots actually pick from a bin? The measured rates

A state-of-the-art robot bin-picking system succeeded on 82.3% of its real-world picks, and only 73.3% on hard objects. Robots do not pick perfectly, and the failure rate rises with difficulty.

A separate system, tuned for speed on densely packed metal parts, reached up to 600 picks an hour at 96 to 99% grasp success. Together the two studies show what bin picking actually delivers on real hardware, as opposed to in a demo.

This page traces both sets of measured rates to their studies and separates success rate from throughput, two different things that vendor claims blur.

Data covers Measured bin-picking studies on real robot hardware (2024 to 2026). Last reviewed by a human editor before publication.

The figures and where they come from

Each figure is rated for how safely you can cite it today. Ratings judge current usability, not whether a number was ever correct.

Figure	What it is	Source	Citation Confidence	Notes
82.3%	Overall grasp success (suction, real hardware)	[A]	Medium	Real-world grasp success for a state-of-the-art suction system, 176 of 215 attempts. About one pick in six failed.
73.3%	Grasp success on hard objects	[A]	Medium	Success dropped to 73.3% on hard objects, from 90.9% on easy ones. Difficulty drives the failure rate.
up to 600 picks/hour	Throughput on dense metal parts	[B]	Medium	A low-cost bin-picking system reached up to 600 mean picks per hour on densely packed metal parts.
615.89 picks/hour	Best measured throughput	[B]	Low	The best single-object throughput measured, at 96 to 99% grasp success. Throughput and success are reported together.

Why the numbers disagree

Bin-picking demos show flawless runs; measured studies show a failure rate. On real hardware, a leading suction-based system succeeded on 82.3% of picks, meaning about one in six failed, and success fell to 73.3% on hard objects. The gap between a curated demo and a measured rate is the story.

Success rate and throughput are also different metrics that get quoted interchangeably. One study reports a success percentage; another reports up to 600 picks an hour at 96 to 99% success on easier, densely packed parts. A high picks-per-hour figure on easy objects is not comparable to a success rate on hard ones, so pairing the two without context misleads.

The results depend heavily on the objects. Easy, rigid, well-separated parts pick fast and reliably; irregular, deformable, or tangled objects drop both success and speed. A single bin-picking number without the object set behind it says little about a different application.

How to cite these figures

Separate success rate from throughput. Cite 82.3% overall success (73.3% on hard objects) from one study, and up to 600 picks an hour at 96 to 99% on dense metal from another, without merging them.

Always attach the object difficulty. Bin-picking rates swing with the parts; an easy-object figure does not predict hard-object performance.

Treat vendor demo rates skeptically. Measured real-world success runs below the flawless demo, and one pick in six can fail on hard objects.

Where people go wrong

Quoting a throughput figure as if it were a success rate, or vice versa. They measure different things and often on different object sets.

Assuming a bin-picking rate transfers across object types. Success and speed both fall on hard, irregular, or tangled parts.

Taking a demo success rate as typical. Measured real-world rates are lower, with meaningful failure on difficult objects.

How we checked

The figures come from two open-access robotics papers reporting real-hardware bin-picking results. We retrieved both from arXiv and confirmed the 82.3% and 73.3% success figures in the first and the 600 picks-per-hour and 615.89 throughput figures in the second.

We keep the two studies distinct because they measure different things, success rate versus throughput, on different object sets. Merging them into a single bin-picking number would be exactly the error the page warns against.

Both are single studies with specific hardware and objects, so we frame the numbers as measured examples of what bin picking delivers, not as universal rates, and we tie each figure to its object difficulty.

Full source list

Primary sources, with live links. Every figure above traces to one of these.

[A]arXiv preprint2024
"OptiGrasp: Optimized Grasp Pose Detection Using RGB Images for Warehouse Picking Robots" (arXiv:2409.19494)
https://arxiv.org/html/2409.19494
[B]arXiv preprint2026
"Pickalo: Leveraging 6D Pose Estimation for Low-Cost Industrial Bin Picking" (arXiv:2604.04690)
https://arxiv.org/html/2604.04690v1

Common questions

How reliably do robots pick from a bin?: Not perfectly. A state-of-the-art suction system succeeded on 82.3% of real-world picks, about one in six failing, and only 73.3% on hard objects. Measured rates run below flawless demos.
How fast can a bin-picking robot go?: A low-cost system reached up to 600 picks an hour, best case about 616, at 96 to 99% grasp success on densely packed metal parts. Speed and success depend heavily on the objects.
Why do the two studies give such different numbers?: They measure different things, success rate versus throughput, on different object sets. A high picks-per-hour figure on easy parts is not comparable to a success rate on hard ones.
Do these rates apply to my parts?: Only as a guide. Bin-picking success and speed swing with object difficulty, so easy, rigid, separated parts perform far better than irregular, deformable, or tangled ones.

More data, traced to source

A robot's repeatability is not its accuracy: the measured gap
A robot's repeatability spec is a sub-millimeter number; its actual accuracy out of the box is far worse. In one measured study a high-precision arm was off by 1.7730 mm before calibration, and calibration cut that to 0.1041 mm.
A cobot's measured productivity gain: 10%, not a multiple
Collaborative-robot marketing implies large throughput gains. A time-studied assembly cell measured the actual gain at 10%. Here is the study, and why a measured number beats a marketed one.
Robot reliability numbers: the vendor claims and the one independent study
Manufacturers advertise robot uptime in the high nineties and mean time between failures in the tens of thousands of hours. The one independent study of more than 400 factories found a robot cell is reliable 88 percent of the time, with 87 minutes between failures.