-
Why a tuned drone stops 5 cm short of its set point: the integrator's slow pole, and a fix with no crash cost
One in four auto-tuned simulated quadrotors ends a 1 m step 5-8 cm short. A three-state position-loop model predicts which; gating the integrator fixes it.
-
Widening a grid repair's trust box fixes its empty QP, then makes it repeat the step it just rejected
Re-solving an empty repair QP at the smallest workable trust box fixed all 94 empty QPs, but 68 of them were the QP just rejected. Net: fewer finishers.
-
216 research figures redrawn so every legend label is on the page
A legend fix in the shared figure helper only helps figures drawn after it. Redrawing the old ones put 230 lost labels back, with no data point moved.
-
A battery drift forecast that got closer and made the plating forecast worse
Tying the plating-overpotential drift to the charge overpotential cut its ten-cycle error by up to 73%, and still made cycle-99 lithium-loss forecasts worse.
-
A lithium-plating forecast trick, tested on 32 battery designs it never saw
Anode node-gain profiles learned on 40 simulated cells still help a ten-cycle plating forecast on 32 new ones, but by less, and with big single-design misses.
-
A data-sheet stand-in for the gear-pump sizes an optimiser picks closes a fifth of the gap
Tilting the published pump-size weight by each gear pump's own idle drag gets the one-number excavator rule to 0.0053, against 0.0056 and an oracle's 0.0042.
-
A free gyro time offset fixes the focal length, but the offset it reports is 3 ms off
In a 1.2 s self-calibration window, a free gyro-camera offset cuts a late gyro's focal error from 0.81% to 0.05%, but the offset absorbs unmodelled translation.
-
Calibrating each anode node's lag from ten cycles helps one battery set and not the one that needed it
A per-node gain fitted on cycles 3 to 10 takes Chen2020's cycle-99 plating forecast from 5.4% to 3.9% but leaves OKane2022 at 6.9% against 6.7%.
-
A gear-pump pick table with synthetic bend rows beats the two-price idle rule, by a little
Rows built by bending donor pumps' idle curves let a builder-side pick table beat the two-price gear-pump rule, 0.0052 against 0.0056 left-out RMS.
-
A published pick table cannot stand in for a gear pump's own picks, because one pump bends alone
A builder-side pick table makes the one-number excavator gear-pump idle rule worse (0.0075 against 0.0056): no table row bends like the Danfoss family.
-
A research page 427,675 pixels tall, folded to 11,000
The research programme pages printed every log entry open. On a phone the median was 64,780 px. Folding older entries brings it to 9,470.
-
A solver's evaluation count is not a reproducible number, so a test should not pin it
A research test failed locally, 42 > 22. The stored count came from another SciPy build; about half of 41 counts matched on rerun, every f agreed.
-
A wrong readout time costs a self-calibrating camera more than a misaligned or late gyro
A 10° camera-IMU rotation adds 0.42% to a window's focal error and a 10 ms lag 0.35%, but a readout 5% off adds 1.23%, always the same way.
-
How far apart should delivery-drone altitude layers be? Seven milestones, one answer: about 44 ft
A paper draft from seven aerolog-scaling milestones: with metre-scale altitude error, drone layers under 400 ft need 13.3 m spacing, and drift is mostly shared.
-
Real drones hold altitude to 4 cm. The barometer, not the controller, decides how close layers can be
17 public PX4 flight logs: the altitude controller tracks its estimate to 4 cm RMS. Barometer error, up to about 1 m, decides between 7 and 5 drone layers.
-
Against RTK height, a drone barometer errs by about 0.8 m in forward flight, and the layer stack keeps 5 bands
400 public PX4 logs, 98 with an RTK fix: barometer minus RTK height is a median 0.83 m RMS in level forward flight and 0.35 m in hover.
-
Barometer drift since take-off: nearby drones share half of it, so altitude layers can stay 13 m apart
NOAA pressure data: drones zeroed at take-off share most of their barometric drift, so a 400 ft stack keeps 5 layers. Treating it as independent costs 12%.
-
An error budget for a ten-cycle battery plating forecast: the anode OCP, not the capacity fade
Scoring all eight forecast-or-measured input combinations on 40 PyBaMM designs shows where a ten-cycle lithium-plating forecast loses accuracy on LG M50 cells.
-
Audio sync of cameras under typing noise: flatten the loudness first, and 78% wrong becomes 0.1%
Dividing each device's audio by its own 32 ms loudness envelope before cross-correlating fixed sync under keyboard, door and chewing noise in 62,000 windows.
-
A plating forecast can borrow how anode layers lag from other designs, on one cell chemistry out of four
A node-gain profile carried from sibling designs cuts OKane2022's ten-cycle plating forecast error from 6.7% to 4.1%, but worsens two other parameter sets.
-
Two more test years rescue a permit-forecast selector: 2019 and 2020 were the odd ones out
On 2017 and 2018 the 3-year staleness cut wins again and the six-year-old-permit selector picks it. Pooled over 12 city-years it beats every fixed rule.
-
No permit staleness cut wins every year, but a 4-year cut is never the worst
Over 12 city-years, dropping source permits open past 4 years leaves 0.78 days of q0.9 regret, against 1.67 for a 5-year cut and 1.73 for a 3-year cut.
-
Audio sync of cameras: a smart switch between two correlators did no better than always using one
A per-window switch on frame-energy kurtosis matched always-normalised GCC (11.5-12.5% of windows a frame off, plain GCC 33-35%) but was no better.
-
Audio sync of phones: a simulated room picked the right fix, but guessed the failure rate low
On five real phones a simulated room ranked ten noise types right (Spearman 0.89-1.00) but ran low: white noise failed 5-10% real against under 1% simulated.
-
A battery's charge-capacity fade follows its electrode balance, but that doesn't fix the plating forecast
The electrode-balance window forecasts cycle-99 charge capacity within 4 to 8% from ten cycles, but the plating forecast's error stays: it is the lever.
-
A small local language model knows which way religious groups differ, but not by how much
Qwen2.5-1.5B on a CPU, given only a GSS respondent's background, ranks unasked items at AUC 0.57 and predicts group gaps at a tenth of their size.
-
DiLoCo compression at H = 32: between the H = 16 tie and the H = 64 gain, and within noise of zero
With a dense last sync, rand-k keep 1/4 ends 0.006 nats below uncompressed DiLoCo at H = 32 (se 0.008): between H = 16's tie and H = 64's 0.018 gain.
-
Don't tune a drone for a slower motor than you measured: train across the range instead
In simulation, a quadrotor tuned at its measured 39.1 ms motor lag but trained up to 150 ms crashes as rarely as a 75 ms tune, and 5x less than a narrow tune.
-
False ties that close triangles hurt village seeding more than random ones
On 43 village networks, survey respondents naming false ties leave the seeding advice intact, but a false friend-of-a-friend costs more than a random one.
-
Keeping warp and weft apart made our cloth simulator's drape error worse, not better
An orthotropic cloth element that keeps warp/weft bending and stretch apart over-drapes nine measured wovens slightly more than the isotropic one: RMSE 0.161.
-
How wide a motor-lag range should you train a drone controller over? At least 4x what you measured
In simulation, training to 100 ms of motor lag left 3x the crashes of 150 ms; 200 ms cost nothing at the measured lag and cut slow-motor crashes by 2x.
-
The nailing humanoid's hip joint was held on the wrong side, not short of sensor
Flipping one held hip joint's side lets the H1's worst nailing pose take the full preload: 0.172 mm falls to 0.005 mm with an FT 300-class wrist sensor.
-
One hip joint sets the wrist sensor a nailing humanoid needs
A hold-first preload can't rescue the H1's worst nailing pose: the stance caps one hip roll at 0.034 N m, so it needs a sensor with under 0.19 N z error.
-
Choosing a humanoid's hold sides per pose helped the H1 and pushed the G1 over spec
Per-pose hold sides cut the H1's worst nailing residual at 20 N hose drag from 0.032 to 0.010 mm, but raised the G1's from 0.011 to 0.037 mm.
-
Scoring a humanoid's hold sides under hose drag fixed the G1 and exposed a coverage gap
Hold sides chosen over 17 hose loads return the G1 to 0.011 mm at 20 N drag, but 17% of its swept cases have no plan at all.
-
How much air-hose drag can a nailing humanoid take? At a 1.2 m grip, about 4 N
No forward lean keeps a humanoid nailing high on a wall balanced under 20 N of air-hose drag. At a 1.2 m grip it tolerates about 4 N pulling back.
-
Not a solver bug: why a nailing humanoid tips back when its air hose pulls
Eight robot poses had no foot-force plan under a backward air hose. We checked the solvers: the robot was tipping back, and the hose ate 15 N of headroom.
-
Half of a battery plating forecast's error is the lever, half is the law's straight line
Given the true anode potential rise, a ten-cycle lithium-plating forecast's error on LG M50 cells halves, from 8.2% to 3.8%. The law's line holds the rest.
-
Underwater glider or propeller AUV? The answer is one ratio, and it crosses at about 4.9
Over 2,430 modelled cases a buoyancy glider out-ranges a propeller AUV once the AUV hotel load passes about 4.9 times its own.
-
Which hull carries a survey payload furthest per kWh? The marine-hull result, now a paper draft
The marine-hull ranking of seven survey-craft body types, set as an IEEE conference draft with each of its 53 quoted figures checked against its results file.
-
Why a hydrofoil survey boat gets a seventh of a monohull's range: the speed it is forced to fly at
Range against speed for five small survey-boat hull types on one 10 kWh basis: the monohull peaks at 824 km near 1 m/s, the foil cannot fly below 9 m/s.
-
The solar boat's best sun-correction clamp is ±2% whether its climatology is current or a decade old
Rescoring clamps from ±1% to ±10% under a current climatology: ±2% stays best and beats no correction in all 11 years, while wide clamps lose their edge.
-
The solar boat's ±3% sun-correction clamp is a plateau, not a lucky pick, and ±2% is the safer end
Fixed clamps from ±1% to ±10% scored on six years no width was tuned on: ±2% to ±4% form a plateau, and only ±2% and ±3% beat no correction every year.
-
At high latitude the solar boat's sun correction only ties the plain climatology, and wide clamps lose
At nine high-latitude points with a fixed pool, a ±2% clamp ties no correction and ±10% loses in five of six years; the old wide-clamp win came from drift sign.
-
Forcing the solar boat's sun forecast darker or brighter: a cautious boat loses more than a bold one
Shifting the learned sunlight ratios by -3% to +3% at nine high-latitude points: every clamp gains brighter, loses darker; an old reading is withdrawn.
-
Most of the solar boat's 2025 sun-correction lead was a decade offset; only the ±3% clamp kept it
Rescoring 2025 against a 2005 to 2014 sun climatology, the learned corrections' lead mostly vanished. Only a ±3% clamp stayed ahead of no correction.
-
The solar boat's one wide-clamp win survives a forced sun bias, so it was never about the bias
Forcing the learned sunlight ratios ±3% in the design where ±10% beat ±2% at high latitude never removes its lead; ±2% still beats no correction every year.
-
Between 70% and 90% tie recall, a village survey's seeding advice only gets better
Ten pooled Monte Carlo copies on 43 village networks: census rules barely care about recall, the half survey gains 0.2 point per tenth, and a dip was noise.
-
A nailing humanoid's hose-drag limits are necessary, but its calibration pose is the weak link
Past each pose's drag limit no nailing plan exists, but inside it 75 of the G1's 96 failures come from one calibration pose under sideways hose drag.
-
Matching a plating forecast to its nearest battery designs did not help; anode thickness might
Borrowing anode node gains from the three nearest battery designs did not beat a set median. A line in anode thickness did, on two of four sets.