AI and Robot Learning
Robot Grasping: What Success Rate Is State of the Art?
Published grasp success rates of 90 percent and above are real and misleading. What the number means, why a 95 percent grasp rate can still fail a production line, and what to demand instead.
Reported grasp success on unknown objects sits around 85 % to 95 % in controlled evaluations, and that is genuinely good research. It is not good enough for production. A cell running 1,200 picks per shift at 95 % produces 60 failures, and the number that matters is not the grasp rate but what happens after each of those 60.
The arithmetic of a good percentage
| Success rate | Failures per shift | Minutes lost at 45 s each | Practical meaning |
|---|---|---|---|
| 90 % | 120 | 90 | Needs an operator present continuously |
| 95 % | 60 | 45 | Operator attends every few minutes |
| 98 % | 24 | 18 | Workable with good recovery behaviour |
| 99 % | 12 | 9 | Feasible unattended for part of a shift |
| 99.5 % | 6 | 4.5 | Practical lights-out threshold |
| 99.9 % | 1.2 | 0.9 | Comparable to conventional automation |
What the published number actually measures
Grasp success in a paper usually means the object was lifted and held for a defined period. Production needs more than that, and each additional requirement lowers the number.
- Lifted and held. The published metric.
- Held with a known pose in the gripper. Required for placement into a fixture, and much harder.
- Held without damage. Relevant for food, electronics and finished surfaces.
- Placed correctly. The end-to-end metric a line actually cares about.
- Placed within cycle time. A grasp on the fourth attempt succeeded and missed the takt.
End-to-end place success is commonly 5 to 15 percentage points below reported grasp success on the same system, and it is the only figure worth quoting in a specification.
What actually moves the rate
| Factor | Effect on success | Why |
|---|---|---|
| Objects touching or overlapping | -5 to -20 pp | Segmentation ambiguity, collisions during approach |
| Shiny or transparent surfaces | -10 to -40 pp | Depth sensors return noise or nothing |
| Deep bin walls | -5 to -15 pp | Approach angles restricted, collisions |
| Deformable items | -10 to -30 pp | Shape changes between perception and contact |
| Very small or very large parts | -5 to -20 pp | Gripper geometry mismatch |
| Known object set with CAD | +5 to +15 pp | Model-based pose estimation is far more reliable |
| Form-fit rather than friction grasp | +5 to +15 pp | Removes friction uncertainty entirely |
| Better lighting and a matched sensor | +5 to +20 pp | Most failures are perception failures |
The bottom three rows are the actionable ones, and they are all engineering rather than machine learning. Reducing the problem, by presenting parts better, choosing an appropriate sensor and designing a gripper that captures a feature, moves the number more reliably than a better policy does.
Detecting the failure is the hard part
A system cannot recover from a failure it does not notice. Four detection methods are used in practice, and most cells need at least two.
| Method | Detection rate | Added cycle time | Misses |
|---|---|---|---|
| Gripper position feedback | 70 to 90 % | 0 ms | Wrong object, partial grip |
| Vacuum pressure switch | 85 to 95 % | 0 to 20 ms | Partial seal that later fails |
| Weight check at the flange | 90 to 98 % | 100 to 400 ms | Two objects instead of one |
| Camera check after lift | 95 to 99 % | 150 to 600 ms | Occluded pose errors |
| Light barrier at the transfer point | 98 to 99.5 % | 0 ms | Nothing, but detects late |
Combining a cheap immediate check with a slower confirming one is the usual answer: gripper feedback catches the obvious failures in zero time, and a camera or weight check catches the rest before the part reaches a fixture where a wrong pose would cause damage.
How to write the requirement
A useful specification contains three numbers and one behaviour, and no percentage on its own.
- End-to-end place success across a defined object set and presentation condition.
- Cycle time at that success rate, including retries.
- Mean time between interventions requiring a human.
- Defined failure behaviour: detect, release safely, retry a bounded number of times, then signal and stop in a known state.
Frequently asked questions
What grasp success rate is state of the art?
Around 85 % to 95 % on unknown objects in controlled evaluations. End-to-end place success on the same system is typically 5 to 15 percentage points lower, and that is the figure a production line experiences.
Is 95 percent good enough for production?
Rarely. At 1,200 picks per shift it produces 60 failures. Unattended operation generally needs 99.5 % or better, unless the system detects and recovers from every failure automatically.
Why do shiny and transparent objects fail?
Because depth sensors rely on returned light. Specular surfaces reflect the pattern away and transparent ones let it pass, so the sensor reports noise or nothing. Success rates commonly drop 10 to 40 percentage points on such items.
What improves the rate fastest?
Engineering rather than learning: better part presentation, a sensor matched to the surface, appropriate lighting, and a gripper that captures a feature instead of relying on friction. Each of these can add 5 to 20 percentage points.
What should a specification demand?
End-to-end place success on a defined object set, cycle time at that rate including retries, mean time between human interventions, and a defined failure behaviour. A grasp percentage alone is not a specification.
Sources
- arXiv robotics preprints, robotic graspingPrimary literature on grasp evaluation protocols and reported success rates
- Robotics at NISTNational Institute of Standards and Technology, robot performance measurement and test methods
- Open X-Embodiment: robotic learning datasetsEvaluation protocols used across the pooled datasets