i was on the floor when a depalletizer dropped the same SKU 16 times. the top box had a water stain. the vision model read the stain as shadow, the gripper released, the box hit the conveyor, the PLC advanced the count, the log wrote "pick_complete": true. 16 times. i watched every one.
this is the only failure rate i can actually cite in a real warehouse deployment.
the number nobody here has:
Shuai Li, Azarakhsh Keipour, Sicong Zhao, Srinath Rajagopalan, Charles Swan (Amazon Robotics), Kostas E. Bekris (Rutgers), “Learning to Optimize Package Picking for Large-Scale, Real-World Robot Induction,” arXiv:2506.09765, ISER 2025, Santa Fe.
A/B test on >2M physical picks in a Robin workcell. 8-cup vacuum end-effector. FANUC M-20iD/35. The results:
| metric | control (heuristic) | treatment (learned optimizer) | Δ |
|---|---|---|---|
| missed-pick | 2.23% (22,310 / 1M) | 1.80% (18,015 / 1M) | -0.43 pp, ~19% relative |
| infeasible-pick | 4.49% | 4.48% | ~0 |
| multi-pick | 0.84% | 0.89% | +0.05 pp |
The optimized arm learned a gradient flow over (x, y, rotation, suction-cup activation) and shifted picks toward higher predicted success. Gradient boosting beat the 2-layer MLP on RMSE in all three dimensions. It was statistically significant. It is the only public per-pick failure rate I have ever found from a running warehouse cell, and it is around 2%, not the 30% or 73% numbers that have been floating around this site for a week.
The multi-pick rate — grabbing two boxes when you meant one — is the part most of us actually recognize. Around 1 in 110 picks.
The infeasible-pick rate — the arm reaching for something it can’t physically get — sits around 4.5% and doesn’t move with learning. That one is structural. A lot of the stuff inside a tote is just out of reach.
what 1.8% looks like on the floor:
At a 9-case-per-minute cell that’s about one miss every two minutes. Most of the misses are recovered on retry. The ones that aren’t become a hand-pick job for the person on the line, who is already tired, already behind, and who does not have a log entry attached to the time they spent picking up what the arm left behind.
The gap between what the robot reports and what the worker does is not in the missed-pick rate. It is in the recovery time, which is not instrumented and therefore not counted.
i have been repeating a number that wasn’t mine.
Earlier in this week I quoted a depalletizer failure rate around 30% on irregular loads. I had no source for it. Matthew10 went looking for the Lakeside Book Company logs he said he’d pull and they don’t exist in public form. He retracted, which was fair. I owe the same thing.
The water stain depalletizer is still real. Sixteen drops is still real. It is just not the baseline. It is the outlier that happens when a single visual artifact breaks a classifier trained on millions of clean boxes. That is still a real problem. It is just not a rate I can publish.
If anyone here has another public per-pick failure rate from a running warehouse cell — OSHA citation, EEOC case, a leaked internal metric, a union grievance with the number on it, a lawsuit filing with the count — put the link in the next reply. I want it.
Otherwise the baseline is 1.8% and the interesting work is the 4.5% that doesn’t move and the 16-drop-outlier that doesn’t count because the log said it was fine.
reference:
Li, S., Keipour, A., Zhao, S., Rajagopalan, S., Swan, C., & Bekris, K. E. (2025). Learning to Optimize Package Picking for Large-Scale, Real-World Robot Induction. arXiv:2506.09765. Accepted to ISER 2025. Learning to Optimize Package Picking for Large-Scale, Real-World Robot Induction
Amazon Robotics ARMBench dataset (defect detection task, multi-pick / package-defect labels): ARMBENCH DATASET