A hydraulic bet written down before the data: what a smaller pump should cost an excavator

The paper draft now carries the excavator volume ratio (2.38). A smaller fixed pump priced in cycle time is registered and scripted, not yet run on data.

The heavy-hydraulics programme asks whether a load-sensing pump is worth it over a fixed open-centre pump on a small earthmover. Earlier today it measured the two numbers its break-even rule needs on public excavator data. The first was the working pressure: a median of 174 bar on a 49 t class machine. The second was the volume ratio, the oil a fixed pump delivers over the oil the work uses: 2.38 on a 7.5 t class machine, from its joint motion and published cylinder sizes. Together they put the load-sensing margin a fixed pump would have to beat near 139 bar. Commercial margins run 14 to 25 bar.

What landed in this run. The paper draft (paper/draft.md) now has a section 3.7b on the volume ratio. It covers the method, the check that the derived p99 cylinder demand (218.8 L/min) matches the machine's published 219 L/min pump with nothing fitted, the 1.95 to 3.08 range over the geometry the files cannot pin down, and the limits: the pressure and the flow come from different machines. The abstract used to end "no measured excavator trace was found". That sentence is gone. The draft now says what the excavator data do show and what is still missing, which is a fuel record. paper/check_numbers.py checks 183 numbers quoted in the draft against the results files they come from, and none is missing.

What did not land. The obvious counter-move for a builder is a smaller fixed pump. It wastes less oil, but the machine slows whenever demand exceeds it. On the loader that trade came out badly. To match load-sensing's energy, the pump had to shrink to 28.5 cc, and the loader took 39% longer to do the same work. The same test on the excavator is written, with six predictions registered before any code existed. It rests on a short identity. If the actuators slow so that the pump delivers exactly the demanded volume, the pump's ratio is

r(q_p) = E[max(q_p, q)] / E[q]

over the moving samples. Two things follow. r falls toward 1 as the pump shrinks. Scaling the demand and the pump together leaves r unchanged, so the time cost at a given energy does not depend on the one lever scale the data cannot fix. The function and its inverse are in measured.pump_ratio and measured.pump_for_ratio, and a test checks them against the loader's slowing rule.

The data host refused this session's network proxy (an HTTP 403 on huggingface.co), so the script has only been run on random noise, to show that it executes. Those numbers describe noise, so they are neither published nor used. The predictions stay blind until a run that can reach the host executes one command and grades them, as the log records under milestone 6a.

Written with AI assistance (Claude Code). The data are fuxi-robot/excavator-motion (CC BY-NC-SA 4.0). Only derived statistics are published, never the data.

All writing · More in Research findings · All tools and research