the protocol

How to enter the board.

Milkbench scores one thing: how long a humanoid hand takes to move a litre. Any hand qualifies, including hands that are not humanoid, so long as the run follows the rig and the trajectory below and reports the four numbers.

1 · The rig

The hand is held at a fixed pose against a compliant artificial teat mounted on a reservoir. No approach, no teat search, no visual servoing. Milkbench measures the milking motion and nothing around it, because the approach problem is already solved by commercial systems and the motion is not.

Teat length, base to tip: 80 to 90 mm
Teat diameter at the base: 22 mm
Teat diameter at the tip: 18 mm
Teat wall: silicone, shore A 20 to 30
Reservoir head above the teat base: 120 mm, held constant
Fluid: water at 20 °C, or milk
Hand pose: fixed for the whole run

Report any deviation from this rig with the run. A teat that is softer or a head that is higher both flatter the result, so a run that changes the rig is listed with the change stated rather than quietly ranked against runs that did not.

2 · The trajectory

Open loop, scripted, and swept. Hand milking has two phases and both have to be present:

Occlusion. The thumb and index close on the base of the teat and seal it, trapping what is in the cistern so it cannot return to the udder.
Sequential distal squeeze. The remaining digits close in sequence, running a proximal to distal wave that drives the trapped fluid out of the tip.

A whole-hand squeeze is not a milking motion and does not qualify. Without the seal the fluid goes back up, and without the wave nothing drives it down.

Sweep two parameters and report the best cell: cycle frequency from 0.3 to 2.5 Hz, and closure amplitude from 30 to 100 percent of the hand's travel.

3 · The four numbers

Volume per cycle: ml
Volumetric flow: ml / min
Cycle success rate: percent
Peak servo load: percent of rated

A cycle counts as a success when it delivers fluid to the reservoir. Report the rate over at least 200 consecutive cycles, with no cycles discarded.

4 · The score

Minutes per litre follows from the flow, and from nothing else:

min / L = 1000 / flow
flow in ml per minute, measured over the full 200 cycle run

The board is ordered by that number ascending. Entries carry no weighting for cost, mass, autonomy or generality, all of which matter and none of which this board measures. A dedicated installation amortises over one task, whereas a general-purpose platform amortises across many, so the comparison this board invites is deliberately the unflattering one.

5 · What to send

Four things. An entry missing any of them is listed but stays unranked.

Paper. A preprint or a published paper describing the hand and the run.
Video. One continuous unedited take of at least 30 cycles, with the reservoir in frame.
Code. The trajectory and the analysis, in a public repository.
Numbers. The four above, plus the swept grid they came from.

Submit a run Back to the board

Milkbench is a benchmark under construction by Prof. Dr.-Ing. Mark Schutera, DHBW Ravensburg, companion to an abstract submitted to the XXVI International Grassland Congress, Leipzig 2027. The protocol is versioned with the board and will not change under a submitted run.