116: Chapter 116 Non-periodic Embodied Mobile Platform

Early morning on July 1st.

On the computer screen, that 16-second four-grid video was still paused.

This clean internal verification data indicated that the state acquisition link of G-01 was already functional.

Foot contact, body posture, spindle load, and synchronized video frames were finally compressed into the exact same timeline.

But Jiang Lin had no plans to upload it.

The second-stage public video of the Low Entropy Workshop could not simply be a lagging playback.

What it needed to display was a complete platform.

A complete system that bridged the pure Physics characteristics of an aperiodic mechanical structure to multi-dimensional perception of contact states, and further leaped to a safety state machine capable of autonomous fault tolerance and supporting phase re-switching.

In reality, he had only a little over a day left.

At six o'clock tomorrow morning, the ninth transmission would begin.

Jiang Lin sealed the four-grid video into a local folder and renamed it.

[G01_contact_chain_0]

Afterwards, he opened three hard-shell cases placed side by side in the corner.

During previous transmissions, the material ratio inside the cases mostly followed the brutal survival logic of the Wasteland.

High-energy-density compressed food, mineral water, broad-spectrum antibiotics...

Later on, there were also textbooks, literature, second-hand workstations, sensors, low-speed permanent magnet generators, wind-solar controllers, and various measuring tools.

This time, in addition to various supplies, Jiang Lin brought three unclosed problems.

[G-01: Aperiodic Embodied Mobile Platform]

[MPS-Kernel: Proof-Carrying Microkernel Search Framework]

[PFR/Marton: Weakly Constrained Structural Compression Under the Finite Field Model]

The first hard-shell case was the heaviest.

Inside were all the parts related to G-01.

Anodized black aluminum profiles, POM load-bearing skeleton + replaceable polyurethane soles, 7075-T6 aluminum alloy spare link rods, custom springs with different stiffness coefficients, high-frequency quenched pawls and ratchets, miniature cross-roller bearings, FOC brushless driver boards, the STM32H7 main control board acting as the brain, industrial-grade high-precision IMUs, isolated spindle current sampling modules, and global shutter cameras.

As well as several sets of housings for foot contact sensing components that he printed overnight with a 3D printer and had not yet fully finalized.

In a crush-resistant USB flash drive nearby were the dynamic parameters of the A1 regular phase platform, the G-01 aperiodic support parameters, the precise CAD dimension tables of three non-repetitive obstacle plates, and the complete log of the sixteen-second failure link just now.

In reality, he only had time to prove that the acquisition link could work.

But in the Wasteland, he needed to verify the survivability of the entire platform in a harsh Physics environment.

Of course, G-01 was never an isolated robot project.

Jiangs Brick proved how local geometric rules could force out global aperiodic order, and what MPS-Kernel aimed to do was to allow local instruction transformations to generate globally correct and faster microkernels.

And G-01 was the mechanical shooting range where Jiang Lin pushed the exact same set of methods into the real physical world for the first time.

The candidates here were no longer comparator sequences, but springs, link rods, foot materials, contact sensors, and support phases.

The counterexamples here were also no longer incorrect arrangements, but a broken link rod, a drifting timestamp, and a load curve forged by wire harnesses.

What Jiang Lin truly wanted to verify was not just whether G-01 could cross obstacles, but whether MPS's method of generating candidates, eliminating pseudo-solutions, and leaving evidence chains could still hold true when the search object changed from a finite state table to a real mechanical system.

If it could, G-01 would no longer be the second-stage public video of the Low-Intellectual Property Workshop, but the candy gears where MPS extended for the first time from paper proofs, program searches, and finite state tables to bite into the real physical world.

The information density loaded in the second box was very high.

Several enterprise-grade solid-state drives, an offline workstation with all unnecessary peripherals removed, and several printed manuals whose page edges had become frayed from his flipping, covered with various colored annotations.

The bottom-layer architecture of the LLVM compiler, GCC backend optimization rules, complete micro-architecture documentation of x86 and RISC-V instruction sets, execution latency and throughput tables of various instructions under different architectures, the Z3 SMT solver manual for formal proof, the environment configuration of the Coq proof assistant, and a massive and rigorous benchmark framework were all sealed by him into the local Git repository.

The half-hour demonstration in Professor Chen Qiming's office had already proved that MPS-Kernel could run successfully on extremely small problems like sort5.

The zero-one verifier could compress the correctness verification of sort5 down to thirty-two binary inputs.

Graph isomorphism and exact reduction were responsible for merging essentially identical candidate networks and reducing meaningless repetitive searches.

The three-layer architecture could logically and clearly split candidate generation, correctness verification, and hardware cost evaluation.

Even within the deterministic search space of a fixed test machine, it could squeeze out a minimal yet stable throughput advantage on a manually polished sort5 algorithm.

But as Pei Li pointed out incisively, the most core engineering disaster had not been resolved.

Sort5 was too small, and its state space was like a shallow puddle.

Real microkernels and real industrial-grade basic math libraries would not stay at five elements forever.

Once the scale was pushed to sort8, rank8, median9, top-k, or even longer data processing kernels, the state space of their instruction permutation and combination would experience a devastating exponential expansion.

By then, the candidate instructions would no longer obediently grow into standard compare-exchange networks.

In pursuit of ultimate speed, it might be the conditional move instruction cmov, the vectorized blend instruction, the packed comparison instruction pminsd, or even a weird combination of several instructions that seemed completely unrelated in human high-level language logic.

What was even more terrifying was the abyss of micro-architecture.

A piece of code might run fluidly on the Zen 3 architecture, but upon switching to the Skylake architecture, it would immediately degenerate into negative-optimization garbage that dragged down the entire system due to minor cache line conflicts or branch prediction failures.

Therefore, on the ninth Wasteland trip, the task of MPS-Kernel was definitely not to make another sort5 with better benchmark scores.

Instead, it was to build heavy engineering machinery to climb over the tall wall pointed out by Pei Li, named exponential expansion and hardware noise.

The third category of items was the smallest.

Namely, hard drives, USB flash drives, electronic ink screen readers, and various writing tablets.

They were packed with textbooks and mathematical literature.

The direction Professor Han Yanshan threw at him during the special report that day relied neither on any hardware sensors nor on any machined tolerance control.

PFR, Marton, Finite Field Model...

What these words truly collided with Jiangs Brick was the exact same deeper problem.

Under what conditions could local constraints force out a global rigid structure?

But this sentence could not be directly written into any formal paper.

It was too crude, too intuitive, and too much like a romantic literary metaphor, whereas mathematics did not believe in metaphors.

Moreover, the local rules of Jiangs Brick were hard, and those uneven boundaries were like precise gears. As long as a single boundary was spliced incorrectly, the collage structure of the entire space would instantly collapse; this hard rule forced out a macroscopic aperiodic hierarchy.

Meanwhile, the small sumset condition in Additive Combinatorics was soft.

It did not give you any explicit boundary restrictions, but only told you in a statistical sense that after undergoing addition operations, the volume of this set did not undergo the drastic expansion it should have.

What it gave you was an implication of compression.

If one mechanically applied the hard-core logic of handling Jiangs Brick to the soft constraints of PFR, what would be obtained would only be a clumsy analogy that looked pretty but lacked logical rigor and possessed no proving value.

In the ninth Wasteland, what Jiang Lin needed to do was hold a scalpel, strip away this illusory analogical shell bit by bit, peel away the cocoons, and leave behind the mathematics that could truly withstand logical scrutiny and be put into rigorous theorems.

July 2nd, 5:57 AM.

Jiang Lin grasped the ropes binding the three categories of items in his hands.

Exactly six o'clock.

The countdown in the upper right corner of his vision hit zero.

The lights in the room, the sound of the air conditioner, the morning fog of the city, and the four-grid video on the desktop receded at the exact same instant.

The next second.

The sound of the wind poured into his ears.

Jiang Lin stood outside the Stone House, raised his head, and glanced at the off-white sky.

The ninth Wasteland began.

Landing review, stronghold patrols, equipment unboxing, food and water inventory—these boring yet life-or-death procedures, having undergone the baptism of the previous eight times, had become so practiced that they formed muscle memory, requiring no longer word-for-word writing into the work logbook.

Meanwhile, on Jiang Lin's logbook, there were only three lines at this moment.

[One, G-01 whole-machine closed-loop.]

[Two, MPS-Kernel minimum viable toolchain.]

[Three, PFR/Marton Finite Field Model bridge.]

This round, forty extreme years were ultimately compressed into three diagrams.

The first diagram was the contact state diagram of G-01.

For the first three years, Jiang Lin spent almost all his time on the mechanical parts.

In the Wasteland, he had enough time to take apart the exact same structure until it could no longer be dismantled.

The A1 regular phase platform took shape first in the dust and wind of the Wasteland.

Uniform mechanical phase, unified overall stiffness, and fixed, unchanging stride length.

On the standard obstacle plate welded to death by Jiang Lin using scrap steel plates, its performance was extremely clean.

The velocity curve was smooth, the power consumption was in the ideal range, and the changes in the roll and pitch angles of the body posture were much prettier than the early version of G-01.

Jiang Lin kept it as the control group for all subsequent tests.

The logic of engineering told him that without a reliable baseline, all subsequent so-called advantages might simply be survivorship bias brought about by the terrain eccentricity of the test site.

Subsequently, G-01 A was born.

It restored Jiang Lin's proud aperiodic phase, non-uniform stiffness, and unequal stride design, but stripped away all control logic, doing purely state acquisition.

This was the generation of platforms that failed the most disastrously.

The strain gauges of the foot load sensors experienced severe thermal drift and sensing hysteresis.

Even the slightest deviation in the IMU's installation position would allow the resonance of the mechanical structure to introduce massive false vibration noise, turning the attitude data into a complete mess.

Spindle current sampling was severely polluted by the high-frequency PWM switching noise of the motor drive board, and the waveform graph was full of terrifying glitches.

Even industrial-grade global shutter cameras would trigger line jitter and occasional USB frame drops under strong electromagnetic interference, causing frame synchronization to collapse.

Once, Jiang Lin, having pulled three consecutive all-nighters, looked at the almost impeccably smooth foot-end load transfer curve on the offline screen and even thought he had ushered in a breakthrough.

But when he excitedly dismantled the foot-end, reality poured a bucket of cold water on him.

That was not any smooth force transfer at all, but rather a wire harness inside the sensor tightening slightly during movement, imposing a completely false extra physical constraint on the foot-end.

The recording system itself had altered the objective state of the recorded object.

That problem appeared once again.

Starting from this day, Jiang Lin stopped treating G-01 like a machine to be repaired.

He created a new directory in his workstation.

[MPS_Physical_Forge]

Kernel was for small programs.

Forge was for physical structures.

If MPS-Kernel searched for comparator sequences, register scheduling, instruction dependencies, and microarchitecture costs, then what this new directory needed to search for was foot-end materials, elastic beam lengths, sensor embedding depths, IMU installation positions, wire harness routing, spindle sampling frequencies, phase window widths, and pawl release thresholds.

Jiang Lin did not intend to let the machine design the robot for him.

At least not right now.

He was simply going to dismantle G-01 into a set of local candidates that could be recorded, compared, eliminated, and reviewed.

Every failure was no longer just something being broken.

It was a counterexample.

A certain local structure, under a certain terrain, a certain temperature range, and a certain support phase, would push the system into an incorrect global state.

This was the first version of MPS-Forge.

It was still far from being an automatic mechanical designer.

It merely forced Jiang Lin to stop repairing machines by intuition and start examining every local failure mode of a machine, just like auditing the local boundaries of Jiangs Brick.

By the sixth year, G-01 B, having gone through countless overhauls and restarts, began to show a stable dawn of hope.

Jiang Lin abandoned the fragile externally-attached thin-film strain gauge scheme and handcrafted a precise internal elastic beam structure.

The outer layer of the foot-end still used a replaceable polyurethane sole, while retaining the POM load-bearing bracket internally. The force was no longer read directly from the sole surface deformation, but transmitted through a very short rigid aluminum seat into the internal elastic beam, and then read by the sensors deeply embedded within to measure the minor deformation after filtering out high-frequency impacts.

The IMU had its installation base redesigned, fixed onto a short aluminum seat closest to the machine's overall center of mass via custom silicone shock-absorbing balls, isolating the high-frequency resonance of the foot-end.

The spindle current sampling circuit was rewired, incorporating Hall current sensors / isolation amplifiers, differential sampling, RC anti-aliasing filtering, star grounding, and shielded wire harnesses.

The global shutter camera was stripped of any control authority and existed merely as a bystander synchronization verification benchmark.

By the ninth year, G-01 B was finally able to stably output four categories of core data in this brutal environment.

Contact load and direction estimation of the foot-end contact state, Euler angle changes of the fuselage attitude, millisecond-level change curves of the spindle load, and the phase windows of mechanical component operations.

Late one night at the end of that year, Jiang Lin rubbed his frozen hands and sketched out the prototype of the first major blueprint on paper.

The horizontal axis of the chart was absolute time.

The vertical axis was clearly divided into four physical domains: contact, attitude, load, and phase.

From this moment on, every abnormal working condition was no longer a vague statement like "the machine is stuck" or "the motor is burned out".

It was dissected into a chain of events with strict causal relationships.

The true value of this chart was not that it looked pretty.

But that it finally gave failure an indexable shape.

In the past, failed samples in the Wasteland were chunks of iron, broken saw blades, blackened circuit boards, and deformed connecting rods.

Now, for the first time, failure had turned into an indexable data structure.

[Abnormal Foot-End Load Bimodal Peak]

[Yaw Angular Velocity Derivative Step]

[Spindle Current Slope Deviation]

[Phase Window Danger Sector]

[Wire Harness Pseudo-Constraint]

[Sensor Hysteresis]

[Real Jamming]

Jiang Lin labeled every piece of abnormal log, wrote counterexamples for every misjudgment, and left certificates of death for every eliminated structure.

This was not black-box training in the sense of machine learning.

This was evidence chain organization in the sense of MPS.

Where did the candidates come from, why did they fail, under what local conditions did they fail, could the failure be reproduced, and which type of structure should the next round of search avoid.

And the appearance of this chart also meant that after struggling in the Wasteland for nearly ten years, Jiang Lin had, for the first time, forcibly elevated the non-periodic hexapod from a movable mechanical toy into a mobile platform that was fully observable mathematically and physically.

After the tenth year, the true soul, G-01 C, began loading the safety state machine.

The actions the state machine could execute were extremely scarce, numbering only five.

Slow down, release the current mechanical lock, actively unload the power of the abnormally contacting leg, control the fuselage to retreat half a step, and re-enter the next support phase.

This was a hell-level torture of searching for a balance point on a tightrope.

If the judgment threshold was set too low, the platform would be like a frightened bird, frequently triggering false protections just by stepping on a small piece of gravel, and its movement efficiency would be shockingly low.

If the threshold was set too high, by the time the algorithm identified the anomaly, the powerful motor torque would have already made the pawls bite down hard, and the mechanical structure would have suffered irreversible metallic deformation.

If one only looked at the foot-end load, the algorithm would foolishly mistake normal obstacle-crossing impacts for wall-collision anomalies.

If one only looked at the spindle current, the lag in current feedback would make you wait until the motor started smoking before realizing the failure had spread throughout the body.

If one only looked at the IMU's yaw angle, it would be extremely easy to misjudge the natural swaying of the fuselage on rugged terrain as an impending rollover.

In the twenty-third year of the Wasteland, with countless scrapped motors, broken connecting rods, false trigger logs, and counterexample tables piling up the working directory, Jiang Lin finally endured and brought forth the version that could truly work stably in complex terrain.

Multi-signal short-window logical judgment.

It no longer relied on a single threshold, but established a dynamic time window.

Of course, the real difficulty was not writing a state machine.

But which signals were qualified to trigger it?

Jiang Lin had initially tried dozens of seemingly reasonable rules.

Foot-end load exceeding threshold—false trigger.

Spindle current slope anomaly—lag.

IMU yaw angle too large—mistaking normal obstacle crossing for rollover.

Phase window entering the danger zone—unable to distinguish between real jamming and transient impact.

MPS-Forge showed its sharpness for the first time here.

It did not invent wisdom for Jiang Lin.

It merely compressed the normal samples, failure samples, false trigger samples, and real jamming samples accumulated over forty years into a counterexample table, and then forced all candidate rules to undergo interrogation.

Rules looking at only a single signal all died out.

Rules looking only at static thresholds all died out.

Rules unable to explain the source of false triggers all died out.

What remained in the end was not the smartest rule, but the rule least easily pierced through by counterexamples.

Within this extremely short window of a few tens of milliseconds, the foot-end load exhibited abnormal bimodal peaks, the yaw angular velocity derivative showed a step rise, the spindle current slope deviated from the normal model, and the phase window slid into the preset danger sector.

Only when at least three of these four conditions were met simultaneously within the same short window would the safety state machine be forcibly activated to intervene.

In the same year, G-01 C completed the first set of breathtaking and complete control experiments.

On the standard obstacle board initially welded by Jiang Lin, the A1 platform serving as the control group still walked cleanly and neatly, even slightly outperforming in speed.

But on the non-repetitive obstacle field filled with volcanic rock debris and broken rebar, after taking three consecutive steps, minor contact anomaly superpositions appeared within the adjacent support window on the A1 platform, ultimately causing a loss of center of mass balance. The motors roared, triggering protective shutdown, and it ended up upside down with all four legs in the air.

And what about G-01 C?

It was like a silicon-based creature with pain perception and instinct.

The moment its front-right foot stepped on a loose piece of volcanic rock, the millisecond-level data stream rushed into the state machine.

Instantly triggering instructions on the verge of overturning, it smoothly decelerated, decisively unloaded the power of the front-right foot, slightly retreated half a step with a posture that defied human intuition, and then precisely found the next stable support phase, exerted force again, and crossed the obstacle.

In the thirty-seventh year of the Wasteland, after enduring a continuous forty-eight-hour extreme sandstorm disturbance test, the scarred G-01 C stood steadily on the final non-repetitive obstacle board.

Jiang Lin opened that directory in his workstation, which had been renamed three times already.

Initially, it was called [MPS_Physical_Forge].

Later, he shortened it to [MPS-Forge].

And at this moment, the first truly completed chart under the directory was officially named—

[MPS-Forge Experiment-01: G-01 Non-Periodic Embodied Mobile Platform v0.1].

On the left side of the chart was the non-periodic mechanical body, covered in scratches and complex in structure.

On the right was the contact state sensing link as rigorous as a nervous system.

In the center was the core polished over nearly forty years: the multi-signal short-window safety state machine.

At the very bottom of the chart were attached long-term test curves of the non-repetitive obstacle field: pass rate, average pass time, energy consumption per unit distance, number of protection triggers, and damage rate of connecting rods / pawls / foot-end modules per hundred hours.

After returning to reality, Jiang Lin only needed to reproduce the minimal publicly releasable version with the most core and rigorously proven logic. Even if it was just a minimal closed-loop, it would be enough to rip open a gap in the real robotics field.