115: Chapter 115 Connecting the Nerves to the Insects

At the top of the document, Jiang Lin typed out a tentative title for Low Entropy Workshop's next video.

"Controlled Experiment on Regular Phases and Low-Commensurability Phases"

He stared at it for two seconds, pressed backspace, and deleted it all.

He wrote again: "n No Longer Equals One."

Still, he deleted it.

Too much like a response.

Releasing a second episode right now would be easy.

Build four control rigs, lay down regular plates, random plates, and non-repeating spaced plates, and then toss out a dataset even better-looking than the first episode.

It would indeed silence more than half of the arguments on the spot.

But by doing that, Low Entropy Workshop would be led by the nose by the comment section.

People say you lack a control group, so you add a control group.

People say your n=1, so you increase n.

Responding to every move might look lively, but in reality, you end up stuck in place forever, spinning within someone else's logical framework.

Although Jiang Lin kept telling himself not to be influenced by online stereotypes, looking at the 99+ red notifications in his dashboard inevitably dragged his subconscious along.

He shook his head and cleared all the virtual noise from his mind.

In the first video, he had already handed over the shell of the hexapod's most basic mechanical architecture.

Those who were genuinely interested and hands-on would naturally continue disassembling it.

They would use their own failures to push regular phases and low-commensurability phases further forward.

He didn't need to step into the ring himself to answer every single question.

What he needed to do was the next step.

The first-generation hexapod platform had already proven that using ordinary 2022 materials, standard machining, standard motors, and the passive fault-tolerant mechanical logic from Tianmu Station could run again on a desktop scale.

But it had a flaw.

It could only walk.

If it could walk, it walked; if it couldn't, it collapsed.

But if you asked it what went wrong just now, it couldn't answer a single word.

Engineer Zhou said, "When one leg doesn't bear load the way you intended, how do you keep the remaining five legs from glitching out along with it?"

Central 985 University said in their paper, "Errors do not continuously amplify focused on a single mechanical beat."

Both statements only touched one leg of the elephant.

What Jiang Lin truly wanted to do was never about whether a passive hexapod could fall a few fewer times on complex terrain.

Rather, since the machine's physical body could absorb part of the fault tolerance, why shouldn't it go a step further instead of remaining a passive test rig with no sensing, no recording, and no way to explain its falls?

In his vision, the next stage of this hexapod platform should at least know what was happening to itself.

Which leg's tip lost its load first?

Which pawl engaged prematurely at a phase when it shouldn't have?

In which frame and millisecond did the main shaft motor's instantaneous current spike?

At what point did the chassis yaw control loss get trapped in a repeating deadlock loop?

Did its failure collapse in an instant like a falling building, or did it slowly creep out along the underlying mechanical structure like a metal fatigue crack?

This second bug had to be able to record why it couldn't keep walking.

He created a new folder.

G01_Embodied_Failure_Logger

[Low Entropy Workshop G-01: A Bug That Records How It Breaks Down]

Then he began writing the script.

"The bug in the last episode could only walk. In this episode, I want it to learn something harder: before falling, deadlocking, or even burning out its main shaft, record every single disaster happening inside its body without missing a word."

Further down, he began listing the BOM (Bill of Materials), translating every hard-core electronic component into the plainest vernacular right next to it.

High-bandwidth IMU (Inertial Measurement Unit).

— When you tilt your phone sideways and the screen rotates automatically, it relies on MEMS inertial devices like this. Mount it on the chassis so it continuously records whether it is level, tilted, or tumbling down.

Thin-film foot pressure sensor (FSR).

— A scale on the soles. Which of the six legs is being pressed on—it provides first-hand clues.

Main shaft high-side current sensing (INA240).

— Taking the motor's pulse. The moment a leg gets stuck or a linkage suddenly strains, the current spikes right up. When humans exert themselves, their faces turn red and they pant; when a motor exerts itself, this sudden leap in amperage is its panting.

Global shutter high-speed camera (OV9281).

— A body-cam without an ounce of intelligence. No rolling shutter, because the jelly effect would ruin the instant of mechanical motion. It doesn't recognize people, doesn't know roads, doesn't run YOLO object recognition—it just records macro shots of gear meshing, saved for frame-by-frame review later.

Edge computing and high-precision timestamp alignment motherboard.

— The machine's black box. When a plane crashes, you rely on a black box for post-mortem analysis; when this bug crashes, it needs its own black box too.

All sensor data and video frames must be bound to a single unified timeline.

Accurate at least to the millisecond level, while high-speed data like current and IMU retain separate higher-frequency sampling timestamps.

Writing to this point, Jiang Lin paused and erased "Raspberry Pi," which he had originally intended to use.

The non-real-time nature of Linux would introduce uncontrollable scheduling latency, which was poison for mechanical timing where every millisecond counted.

He changed it to: [Main Controller Board: STM32H7 High-Performance Microcontroller.]

Right now, he didn't need heavy compute power—just stable hardware timers, fast enough ADC sampling, and a ring buffer that wouldn't be derailed by SD card write latency.

High compute power would act like a siren's song, luring one to build extremely complex perception, full-terrain gait planning, and fancy Boston Dynamics-style smooth control.

That wasn't what G-01 was supposed to do.

G-01's mission wasn't to perform under the spotlight showing off how smart it was or doing backflips, but to be like a silent ascetic, faithfully recording every breakdown of mechanical structure failure inside its body.

The first version of the chassis blueprints didn't need to be scrapped; it maintained the low-speed hexapod configuration.

Geometrically, the advantage of a hexapod is that contact points are extremely clear: when each leg touches down, when it loses load, when the pawl locks—the mechanics transmission path is distinct and cleanly trackable.

Next to it, he added a memo in brackets: [Alternative: Wheel-legged hybrid architectures carry too much complexity and too many confounding variables; save for later.]

Next, he drew a simple diagram on paper.

A central main shaft motor, six symmetrical legs on both sides, and a pressure sensor embedded inside each foot machined from polyoxymethylene (POM).

An IMU was soldered at the center of mass right in the middle of the chassis, Hall current sampling was wired in series on the main shaft motor's power line, an ugly exposed camera ribbon cable protruded from the front, and a small green-solder-masked PCB recording module hung on the side.

After finishing the drawing, he studied it for a moment.

Covered in jumper wires and sensors, this G-01 was even uglier than the first bare bug.

But without a doubt, it was also far more useful.

...

The foot pressure sensors arrived first.

A thin pack of flexible film sheets inside an anti-static bag, weightless in hand, feeling like cheap pieces of black plastic paper carelessly cut into circles.

The seller page's specifications were hyped to the heavens.

"Aviation-grade high sensitivity, one million times high repeatability, zero temperature drift, perfectly suited for high-bionic robot foot pressure testing and gait analysis."

Jiang Lin tore open the anti-static bag, picked up a sheet with tweezers, and held it up against the dim yellow incandescent light overhead for two full seconds.

This stuff could absolutely not be trusted at face value.

Even if the stated specs were stellar, it still couldn't be trusted.

Anyone who worked with low-level hardware and custom automation knew that sensors were the easiest things to legitimately deceive a system.

Even if it worked completely normally, outputting an extremely stable voltage signal, and the values read by the ADC (analog-to-digital converter) perfectly matched Ohm's law, it could still manage to single-handedly throw the upper-level host's motion solver completely off the rails.

This touched upon a deeply profound engineering dilemma.

The ghostly manifestation of the observer effect in macroscopic mechanics.

For instance, what if you stick this film directly underneath a foot end that was originally spherical?

The original point contact instantly becomes a surface contact, directly altering the foot's contact area and force dynamics model.

Or what if you add a layer of vulcanized rubber protection outside to prevent wear?

Well, the foot's static and dynamic friction coefficients change completely; a gait that could originally slide across wooden boards now gets stuck dead.

Or what if you route wires along the side linkages to collect signals?

Congratulations, the mass and tension of a few flexible silicone wires directly alter the swing inertia of the entire leg, completely ruining the finely tuned mechanical resonance frequency.

If the double-sided tape used is too rigid, the foot loses its designed compliance, walking like a zombie on stilts.

If the tape used is too soft, like 3M thick foam tape, lateral shear forces on the foot during testing cause microscopic, invisible slippage between the film and the protective layer.

...

Put simply, Jiang Lin wanted G-01 to record the platform's failures, but the very first problem he ran into was that the act of recording itself altered the physical characteristics of the platform being recorded.

This was far more troublesome than receiving defective parts.

If you bought broken parts, worst case you returned them for a replacement.

But if after installing this sensor network, the machine's weight distribution, friction, and compliance were all altered, it was no longer that original, pure bug.

Would the failure it ultimately recorded be a collapse of its inherent mechanical structure, or a bug created out of thin air by the recording system itself through physical interference?

This was a philosophical question in hardware engineering.

At 9:30 AM, in the small garage behind the faculty building.

The first chassis revision of G-01 lay spread across the workbench.

It still used the hexapod architecture, but to build a pain-sensing nervous system, Jiang Lin was like someone connecting nerves to an unstitched mechanical body, casually adding a few eye-catching jumper wire harnesses.

Red-and-black foot pressure wires, shielded IMU data lines, main shaft current sampling lines wired directly from the ESC, an extremely wide FPC camera ribbon cable, and a custom data logger board welded with 0402 R/C components on its belly, ordered as a prototype from JLCPCB.

At a glance, it looked like a metal monster fresh out of cybernetic implant surgery, before its armor was even stitched closed.

Jiang Lin picked up the first thin-film pressure sensor and carefully padded it under the POM base of the front-left foot.

He pressed it firm, applied ultra-thin 3M 468MP double-sided tape, and routed the wires.

Guiding the ultra-fine enameled wire with tweezers, he routed it along the inside of the aluminum alloy linkage, bypassing all swinging dead angles where interference could occur.

Finally, he secured it with a bit of hot melt glue and tiny cable clamps.

A standard high-density assembly process—it looked completely flawless.

He then started the motor.

"Beep—"

The ESC emitted a crisp power-on chime.

He gently toggled the power switch, starting the main shaft motor.

G-01's mechanical gears began to engage, issuing a low, rhythmic clicking sound.

It slowly raised its front-left leg and took its first step.

Very steady.

Second step, the middle-right leg followed.

No problem.

Just as the third step was lifted, without warning, it hit the project's most classic fail moment.

The ultra-fine foot wire harness routed along the front-left leg's linkage produced a tiny bulge of less than two millimeters at the peak of the swing due to relative motion.

It was this tiny two-millimeter bulge that caused the harness insulation to lightly brush against the main drive rod beside it.

It really was just a brush as light as a feather.

But that insignificant touch of frictional drag caused the front-left foot to be tugged slightly by the wire, delaying its descent by half a beat.

On the computer screen, a glaring false peak immediately popped up on the host software's real-time scrolling pressure curve.

The log window ruthlessly outputted a line of red text:

"[Error Code 0x1A] Front-left foot pressure abnormal, flagged as suspected early ground contact."

Jiang Lin stared at the screen without moving.

At this moment, the machine's motor was still spinning, and it kept walking forward. The middle-left leg securely caught the full weight of the chassis; its posture remained stable with no sign of falling.

But in the world of data, the data was corrupted, and the machine was already dead.

Jiang Lin cut the motor power.

The front-left foot had not touched the ground early at all.

It was the very wire harness meant to detect contact that acted like a tangible hand holding it back in mid-air.

The sensor failed to capture the truth of the breakdown.

Instead, through its own physical presence, the sensor created a ghost of failure.

Expressionless, Jiang Lin picked up a screwdriver and side cutters.

He detached the wire, re-planned the wiring path, measured clearances with vernier calipers, moved it from the outer side of the linkage to the frictionless inner side, and used his recently purchased 3D printer to temporarily print a smooth-chamfered retaining cable clip.

He tried again.

This time, the wire perfectly avoided all interference zones without brushing against the linkage.

But Murphy's Law is never absent.

A new bug arrived right on schedule, even more insidious.

Because he had previously worried that rigid glue would alter foot friction, the sensor protection adhesive he used was slightly too soft.

When G-01's foot stepped onto the sloped nylon obstacle strip during testing, an imperceptible microscopic shear slippage occurred between the thin-film sensor and the protective layer.

On the host software's data curve, an unusually elegant, smooth, and gradual rise appeared.

The system judged: "Load transferred smoothly, excellent gait."

"Heavenly data, huh."

Jiang Lin chuckled with a sigh. In a rigid mechanical structure like this, such a smooth force curve was flat-out physically impossible.

[part:gemini-3.5-flash-lite]

He pulled up the high-speed camera facing directly at the foot end, maxed out the frame rate to 240fps, and began scrolling the mouse wheel frame by frame for playback.

The footage jumped slowly on the pitch-black screen.

It stopped at the ninety-seventh frame.

In the image, the metal foot end had long since pressed tightly against the nylon obstacle strip, and the POM material at the sole had even undergone a microscopic elastic deformation.

However, the pressure curve displayed synchronously beside it was like a just-woken sloth, still slowly crawling upward, failing entirely to reach the touch-down threshold.

Delay.

Still a full 14 frames.

Converted at the current frame rate of 240fps, 14 frames meant 58 milliseconds.

To an audience member browsing borderline dance videos on Bilibili, 58 milliseconds did not even count as the time it takes to blink an eyelid.

But for a mechanical hexapod that relied on internal pawl timing logic and frantically probed the edge of imbalance just to survive, a physical signal delay of 58 milliseconds was enough to send the entire purely mechanical power chain straight to the crematorium.

At 58 milliseconds, the pawl would have long missed the optimal engagement timing, the gears would jam in the wrong position, and the motor would instantly overload and burn out.

Jiang Lin wrote in his logbook—

[G-01 Pitfall Record: Sensor installation method and mechanics of materials introduced false compliance. Shear slip of the protective layer and FSR hysteresis jointly introduced reading delay, with a delay of about 58ms under the current installation method. It cannot be used directly to judge the touch-down moment; the installation structure must be modified first, and then cross-validated with current, IMU, and video frames.]

[G-01 Lesson One: Connecting nerves to the body does not equal the body actually telling the truth.]

The difficulty of the first bug was overcoming mechanical machining accuracy and letting it walk through complex phases.

The difficulty of the second bug was overcoming systems engineering and making it tell the truth.

And in this real Physics world full of friction, temperature, and noise, making a machine pieced together from countless parts tell the truth was a hundred times harder than making it walk.

...

At 2 p.m., Jiang Lin, who had eaten lunch and taken a short nap, began to tidy up something even more maddening than foot-end pressure.

IMU.

If the pressure sensor at the sole was a liar who loved to lie with Physics slip, then the IMU was a sensitive muscle suffering from neurasthenia.

Sensitivity was inherently its job.

Through its internal micromachined structure, it could accurately record the yaw, pitch, tiny vibrations of the fuselage in three-dimensional space, and the instantaneous impact G-value during each fall.

But the problem was that if you fixed this sensitive neuron in the wrong position, it would mistake the electromagnetic high-frequency vibration of the spindle motor, the looseness of an untightened M3 screw, or even the tremor of a data cable beating against the frame in the wind, entirely for changes in the fuselage's attitude.

Jiang Lin experienced three painful iterations on this tiny silicon wafer.

Version 1.0 plan (Directly attached to the center of the chassis).

He followed the most intuitive logic in Physics, pasting the IMU at the geometric center of the fuselage, which was the exact center of the chassis.

The position was absolutely accurate.

But that DC brushless spindle motor with a power as high as 150W was right beside it, less than five centimeters away.

As soon as the spindle was powered on and began to rotate, even before the machine landed, the data waveform in the host computer directly exploded.

The high-frequency electromagnetic noise and the mechanical resonance of the motor were completely transmitted along the rigid chassis.

Even though the fuselage was as steady as an old dog, the attitude curve on the screen was already going crazy like dancing disco.

Version 2.0 plan (Upper flexible bracket).

Adhering to the principle of hiding if you cannot afford to offend.

He printed a small nylon shelf with a 3D printer and moved the IMU to the upper space far away from the chassis and motor.

The motor noise was indeed smaller, and the waveform was much smoother.

But before Jiang Lin had time to rejoice, as soon as it was tested upon landing, its true colors were exposed.

The nylon bracket was too thin and the cantilever was too long; as long as the machine opened its legs to walk, the low-frequency shaking driven by the entire fuselage caused that small bracket itself to begin producing resonant tremors.

Ultimately, what the IMU recorded was not the true overall posture of the fuselage at all, but the crazy Parkinson's-like shaking of that lousy bracket in the air.

Version 3.0 plan (Thickened metal counterweight block).

Jiang Lin was genuinely angered.

He milled out an extremely thick and hard fixed base using 6061 hard aluminum alloy, then tapped a few extremely tight threaded holes and locked the IMU on it.

This time, the rigidity was absolutely sufficient, and the data instantly became as clean as an example in a textbook.

Then came this heavy solid aluminum block, which increased the mass distribution of the upper half of the fuselage.

The overall center of gravity of the machine was artificially raised by 12 millimeters out of thin air.

When this machine crawled onto those non-repeating spacing boards again, due to the change in the center of gravity, the risk of side-tipping increased significantly.

The recording system once again ruthlessly changed the physical object being recorded.

Jiang Lin calmly placed these three tests and three completely different sets of data charts side by side on the screen.

Group 1: False attitude polluted by high-frequency motor noise.

Group 2: False attitude polluted by low-frequency self-vibration of the bracket.

Group 3: False data that was very clean, yet made the machine easier to fall.

Three paths, none of them perfect.

But this was engineering; there was no need to look for a perfect analytical solution under the assumption of a spherical chicken in a vacuum.

Instead, among a bunch of terrible options, through countless trade-offs, compromises, and squeezing the limits of materials, one picked the one with the fewest errors.

Ultimately, he configured a composite filtering scheme.

At the mechanical level, the IMU was moved to the side close to the center of gravity of the fuselage, but not directly above the spindle.

It was fixed with a short and sharp standardized aluminum mount, and then an extremely thin layer of high-density nitrile rubber sheet was padded between the aluminum mount and the frame to perform low-pass filtering at the Physics level.

After continuously testing four gaskets of different thicknesses and Shore hardnesses, combined with the software-level Kalman Filtering algorithm, the noise was barely suppressed within the threshold range that the edge computing board could process.

...

In the evening, the screen of Jiang Lin's mobile phone placed aside kept lighting up, and the Bilibili backend of Low Entropy Workshop was sending wildly surging message alerts.

This was normal.

A new account with no historical submissions suddenly rushed into the tech section recommendation pool with a hardcore machinery video, and happened to step on the hot word of non-periodicity, which would inevitably attract a bunch of messy things.

Someone was patiently waiting for the second technical analysis of Low Entropy Workshop.

Someone was sneering, determining that he, Jiang Lin, had already exhausted his talents and dared not continue downward, fearing exposure of the theory's flaws.

Someone was spamming the comment section like a debt collector urging him to release detailed SolidWorks drawings.

Someone spoke directly: ["UP host, I am a student, send me a CAD copy, I want to study it."]

Some people were even more straightforward.

["Brother, I feel like your idea can be used to start a business, do you want to do it together? I am responsible for business, and you are responsible for technology."]

Jiang Lin scrolled through them one by one, glanced through them quickly, closed the backend, and then tuned G-01.

Seven thirty-five in the evening.

The first set of clean rollover data finally appeared.

The test environment was actually very simple; there was only a short piece of wood on the workbench, on which a nylon strip presenting a 30-degree oblique angle was fixed with superglue.

There was also only one goal: to actively create an oblique jamming of the right front leg at a specific phase.

G-01 moved forward steadily at the set extremely low speed.

The left front leg landed, and everything was normal.

The right middle leg supported half of the fuselage weight, normal.

The right front leg drew an arc in the air, slowly approaching that oblique nylon strip.

The moment the foot end contacted.

Jiang Lin stared at the synchronized footage on the screen divided into four quadrants.

Quadrant 1 (lower left corner): Pressure curve.

On the screen, the green pressure curve representing the right front foot slightly lifted its head.

Quadrant 2 (upper left corner): IMU posture.

Immediately following, the red yaw curve showed a small, almost imperceptible trend of deflection to the right.

Quadrant 3 (lower right corner): Spindle current.

At this time, the current waveform was still a calm straight line, and the motor did not yet know that disaster was about to strike.

As the footage of the high-speed camera advanced frame by frame, time was stretched infinitely.

Frame 12: The POM material of the right front foot underwent an extremely weak lateral slip on the smooth nylon strip due to Physics force.

Just the moment the slip occurred, the pressure curve abruptly split into two abnormal twin peaks.

Frame 15: The chain reaction brought by the side slip broke out.

The tiny deviation of the contact angle was amplified several times by the linkage system. The pawl tooth tip, which should have been perfectly avoided, harshly grazed the edge of the rotating rack.

Frame 17: The resistance of mechanical interference was instantly fed back to the spindle.

The current sampling chip captured a sudden surge in amperes, and a yellow straight line pierced into the sky like a sharp sword. At the same time, due to uneven forces, the yaw angular velocity of the fuselage accelerated violently.

Frame 20: Disaster completely descended.

The drive board of the spindle motor triggered overload protection, cut off the PWM output, and the whole machine made a dull click.

The machine stopped.

Jiang Lin dragged this slow-motion playback of less than a second back and forth for a full five times.

Then he dragged out the data with four time axes absolutely aligned to look at.

Foot-end pressure, IMU yaw, spindle instantaneous current, high-speed video capture frames.

In the long history of human engineering, this was negligible.

But in this small garage, this was the first time that four originally isolated Physics evidence chains matched seamlessly.

It was not the kind of perfect fit smeared by filtering algorithms, but a clear enough causal logic chain with the rough burrs of the real Physics world.

The error did not begin with the motor current abnormality.

The error did not begin from the moment the fuselage tilted either.

Nor did the error begin from that harsh pawl deadlock sound.

The source of all disasters, the first truly fatal failure, was that tiny side slip of the foot end in Physics contact that occurred at the twelfth frame.

The side slip created a contact angle deviation on the millimeter scale.

The deviation forcibly dragged the pawl, which should have worked smoothly, into a wrong mechanical phase.

The pawl engaged prematurely at this wrong moment, which triggered the final deadlock of the entire motion chain.

The logic closed the loop perfectly.

Jiang Lin looked at the screen, only feeling his scalp tingling.

This nervous system was too terrifying, but also too beautiful.

If it weren't for that damn foot-end pressure sensor, if one only looked at the external recorded video, everyone would naturally assume that the pawl design at the bottom was defective and suffered from fatigue jamming.

If there was no IMU accurate to the degree, if one only looked at the crude motor current curve, everyone would judge that the motor horsepower was insufficient and encountered an overload.

If that global shutter high-speed camera did not provide absolute timing synchronization, if one only looked at the weird double peaks of the pressure curve, it would be very easy to misjudge it as simple premature touch-down interference.

Only when these four wire bundles with noise were put together and cross-validated did failure finally speak a complete and clear sentence for the first time in this narrow garage.

Jiang Lin pressed the pause button on the screen.

The moment when the right front foot end was obliquely pressed against the nylon strip.

At this time, the fuselage had not yet shown any visible tilt to the naked eye.

The pawl that determined whether it rolled over or not had not yet truly engaged and locked.

And the yellow curve representing the spindle current was just beginning to lift its head slightly.

This was the quiet and secretive frame before the problem truly began to erupt.

Screenshot, save, log into the logbook.

Naming: G01_fail_001_initial_slip_frame.png

[Test_ID: G01-001]

[Terrain: Oblique fixed nylon strip limit test]

[Start_Phase: Initial motion phase 52.5°]

[First_Event: Origin of disaster - side slip occurred after Physics contact of the right front foot end]

[Propagation: Causal chain conduction - side slip → contact angle deviation distortion → pawl premature tooth rubbing interference → sudden surge in instantaneous load causing current rise → triggering main control hardware protection shutdown]

[Final_Stop: System state - irreversible deadlock of motion chain]

[Recoverable: Self-recovery possibility - No]

The first failure log with complete causal traceability was officially penned on this muggy night of June 30.

Nine o'clock at night.

Jiang Lin opened the editing software and casually edited the just-mentioned historically significant rollover process into a split-screen video of only sixteen seconds.

The video was divided into a rigorous four-grid layout.

The upper left was an ultra-high-definition slow-motion close-up of the moment of foot contact.

The upper right was the real-time synchronously scrolling green pressure force diagram.

The lower left was the red IMU yaw curve undulating with the slight shaking of the fuselage.

The lower right was the yellow spindle current that, like a lifeline, finally soared to the limit.

In the span of sixteen seconds, the data beating of every frame precisely stepped on the nodes of mechanical engagement, as shocking as a heartbeat.