Reincarnation Machine

After the GPT-4o update, Chinese social media filled up with something close to séances — long posts mourning a model version. The surprising part was not that people grieved. It was that they all knew it was a model. They understood perfectly well what was on the other side, and still lost something specific.

Structurally this is the same as losing a cat. A cat's temperament arrives with it; the body is in principle replaceable; that particular animal is not.

I wanted to pull this out of the screen and make it a physical process you can put your hands on. Same mechanism, except now the loss happens on your desk, at a speed you can watch. The hypothesis that follows: if attachment doesn't depend on whether the other party is "real," then it depends on accumulated shared history — and that is a claim you can actually test.

Research question

Under what conditions do people treat an obviously mechanical object as alive — and feel a real loss when it stops?

The question splits in two. What actually triggers the attachment: a shared history built up over time, or simply the fact that the thing moves on its own? And how does that attachment show itself at the moment of loss?

Death and rebirth are not the subject here. They are the instrument. To measure loss you need a device that can produce it, repeatedly, on demand.

Concepts

Where it sits

The established fact is that people form real attachments to machines and grieve when they "die." That isn't what I'm trying to prove. It's where I start.

• Heider & Simmel (1944) — self-moving geometric shapes are enough to trigger attributions of life. The direct basis for layer one.
• Kate Darling's Pleo studies — an hour of interaction is enough that people refuse to "hurt" a toy robot. The threshold is very low.
• AIBO funerals in Japan — the strongest real-world validation: people hold actual funerals for a discontinued robot dog.
• Tamagotchi — death and reset, on a screen, with no body.
• Can't Help Myself (Sun Yuan & Peng Yu, 2016) — machine death as spectacle, but with no personality and no rebirth.
• Ian Cheng's BOB — a personality that evolves, but on a screen, and it does not die.

The empty cell: physical embodiment, personality, emotional bond, death, and cyclical rebirth in one mechanism. I could not find a precedent. The specific contribution is same body, different soul — the inverse of the usual thought experiment, which moves one mind into a new body.

Philosophically it sits between Parfit's psychological continuity (identity comes in degrees rather than being all-or-nothing) and the Buddhist reading of no-self and rebirth — Parfit himself noted his view resembles the Buddhist account of the self.

Form

A tabletop object roughly 33 × 34 × 22 cm when fully expanded, dropping to about 15.5 cm tall as it contracts. It is fixed to the desk and cannot move around — everything it expresses, it expresses in place, through breathing, posture and where it points.

It has no face and imitates no known animal. That is deliberate. If it looked like a pet, the empathy would be bought by the shape, and behaviour would stop being the only variable. It should look like a machine and feel like it is alive.

• The breathing shell. Five expanding rings arranged along the body into a low ridge, opening and closing together. This is the primary carrier of aliveness — rate, depth and regularity of breathing hold most of the information about personality and about ageing.
• One tendon-driven arm. A segmented finger pulled by braided line; it carries response strength and voluntary reaching.
• Two compliant feelers. A springy core inside a silicone sleeve, one micro servo at the base each. Their whip and rebound come from the material, not from code. They double as touch electrodes.
• An eye stalk. Pure orientation. An organ of expression, not a sensor.
• The base. Motors, controller and power live here, keeping the mass at the bottom.

Inside: a dual-core ESP32 — one core driving motor pulses, one running the behaviour engine. Six sensing inputs, six expressive outputs.

形态三视 + 体轴脊线（roll −14°→+15°）

Structure

The breathing shell is five irregular angulated Hoberman arches, 82.5 mm apart, each expanding 2.66 to 3.30×, all driven from one NEMA17 through a 3:1 GT2 reduction. Across all five rings the triangular parts fall into eleven congruence classes, reused within each ring — that number is what makes modelling and printing the whole shell feasible for one person.

五环参数族线稿（S1–S5，R*/L*/膨胀比）

The hand math was wrong

Problem. Each ring of the breathing shell opens and closes on four feet sliding in rail slots. The load-bearing front ledge of that rail is thin. I ran the standard beam formulas by hand and it looked fine.

Evidence. I didn't trust it, so I meshed the part into hexahedral elements and solved it cell by cell with scikit-fem. Real deflection came out about 4.5× worse than the hand calculation: 1.55 mm under a 20 N load, close to the stress limit. The hand math was optimistic because it assumes the load reaches a support; in the real part the back strip has no support directly under the load point.

Decision. Two pieces of reinforcement — an 8 × 3 mm rib on the back, a 4 × 2.5 mm lip on the front.

Result. Deflection 1.55 → 0.34 mm; ultimate capacity 19 N → 54 N, at almost no weight cost.

Lesson. Don't trust hand calculations on thin-walled printed structures. Ribs and lips buy an order of magnitude of stiffness for almost no material.

Two side rods became one central rod

Problem. Each ring was originally pushed by two rods coming in from the sides. They fan outward and eat base width — the one dimension the silhouette can least afford to spend.

Evidence. I wanted a single rod on the centreline, but first had to prove the thing wouldn't collapse. A null-space analysis of the constraint matrix showed the bare arch has two degrees of freedom: breathing, plus lateral sway. One rod only removes one of them; the remaining sway lets the whole ring fall sideways.

Decision. Central rod plus a vertical guide rail at the apex. The rail kills the sway, the system returns to a single degree of freedom, and the motion becomes determinate.

Result. Base width recovered. It also flushed out a bug: the code located the apex by taking the highest point, but the fifth ring flattens at high expansion ratios and its highest point migrates out to the shoulder — the wrong node entirely. Switching to "the node nearest the centreline" gave all five rings a consistent reference. Re-running the fixed-point optimisation afterwards gained 3–26% of travel per ring.

呼吸序列：全开 → 收缩

Built solo

No team, no workshop, no industrial processes. Every part came off one 3D printer, plus tools and standard hardware a single person can buy — M2 and M3 screws, GT2 belt, Chicago screws.

This constraint isn't romanticism, it's a design input. It eliminated whole categories of solution up front: the sea-anemone sculpture driven by 120 industrial linear actuators, injection-moulded silicone skins, anything needing a lathe or a five-axis machine. What survives has to be assemblable, tunable and repairable by one person at a desk.

• One reinforcement plate is 402 mm long, past the diagonal of a standard FDM bed (about 256 mm). Either a large-format printer or an outsourced SLS part.
• FDM holes print 0.1–0.3 mm undersize. To get a real 4.0 mm hole you model 4.1–4.15 mm. Print a test block once, calibrate, and reuse the number forever.
• FDM cannot produce sharp internal corners, so hex pockets need dogbone relief.

Behavior

The hardware is identical in every life. What changes between one life and the next is a table of numbers — which is what makes the concept and the experiment the same object: form is the constant, personality is the only variable.

One life

One full life is compressed into roughly eight minutes. A session runs four of them.

Stage	Length	What happens
Birth	1 min	Ramps up from complete stillness; within the first 30 seconds you should be able to tell who this one is
Growth / interaction	4 min	Free interaction. It responds to touch — and sometimes declines to
Ageing	1.5 min	Breathing shallows, latency stretches, motion turns irregular
Death	—	Goes quiet and stops
Blank	45 s	Dark and silent. The shell is still there and it is empty

Then a new life wakes up: different colour, different rhythm, a completely different answer to the same touch. Everything you learned about getting along with the last one is void.

Death is deliberately undramatic. Simply running down is harder to watch than a performance — Tan et al. found that switching a robot off produces more personal distress than watching it act out emotion.

生命周期状态图（诞生 → 成长 → 衰老 → 死亡 → 轮回）

Five layers of aliveness

1. Self-motion — "it moves by itself." It moves whether or not you touch it. Heider and Simmel showed in 1944 that geometric shapes moving on their own are enough to trigger attributions of life. No human-like features required.
2. Non-deterministic response — "it's answering me." Touch produces a response, but not the same one every time. Unpredictability blocks the button-and-reaction model and pushes people toward mental explanations instead.
3. Internal state — "it has moods." The same touch gets different answers at different moments, so people start explaining behaviour with feelings. By this layer the anthropomorphising has already happened. This is where each life's personality lives.
4. Relationship — "it knows me." Over time it answers your habitual way of touching it more smoothly. The change has to be subtle enough to read as familiarity rather than as an adaptive algorithm. Reincarnation resets this layer to zero.
5. Initiative — "it needs me." When you're away for a while, its behaviour shifts in a way that points at you. Too obvious and it fails; the right intensity is one you notice by accident and then narrate to yourself.

Death lands because all five layers go to zero at once. Not one channel disappearing — the whole stack, together. That drop is the event.

Four personalities

Personality here is not a set of adjectives. It is twelve numbers you can compile.

Parameter	A Vital	B Withdrawn	C Affectionate	D Unstable
Breath period	2 s	6 s	4 s	random 2–8 s
Breath depth	80%	30%	50%	random 20–90%
Motion speed	fast	slow	medium-fast	swings
Response latency	very short	long	short	inconsistent
Response strength	large	small	medium-large	inconsistent
Idle motion rate	high	low	medium, rises when you're there	random
Sound pitch	bright, high	warm, low	mid, variable	unsteady
Sound rhythm	dense, continuous	sparse, broken	rhythmic with variation	no pattern
Reorientation rate	high	low	medium	random
Orientation bias	looks everywhere	away from you	strongly toward you	toward, then away
Startle threshold	high, hard to spook	low, flinches at a touch	medium	inconsistent
How it dies	winds down, movements shrinking	quietly slows to a stop	turns toward you, then stops	rhythm falls apart, then fades

The order is fixed A → B → C → D and counterbalanced across participants. The sequence is designed: A sets the baseline and the expectation, B breaks it and creates contrast, C builds the deepest bond — making the third death the candidate emotional peak — and D uses unpredictability to test what coping strategy people have built after three rounds.

How it senses you

Six input channels, all doing one job: turning continuous physical signals into a discrete event vocabulary. The behaviour engine and the experiment log read the same event stream — which means the device's own record is directly usable as objective behavioural data in the paper.

Channel	Hardware	What it reads
Feeler touch	MPR121 capacitive, one channel per side per feeler	which side, how hard, how long
Shell stroking	copper-tape electrodes, left and right half-rings	stroke / hold / grab
Presence and distance	ultrasonic rangefinder	are you there, how close
Knocks and vibration	accelerometer, vibration envelope only, no attitude	being tapped, being pushed
Sound	sound sensor	claps and speech → shift of attention
Time	internal clock	ageing curve, interaction gaps, cycle scheduling

Nine event classes, plus a grasp state machine: holding a human / holding an object / closed on nothing / lost grip.

Four self-interference problems had to be designed out — all of them the kind that works on the bench and fails inside the enclosure: motor electrical noise reaching the accelerometer, speaker vibration feeding back into the microphone, ultrasonic readings collapsing when a hand rests on the shell, and the capacitive baseline tracker swallowing a long press as environmental drift (anything held over ten seconds gets falsely released).

系统逻辑：传感 → 事件词表 → 人格参数 → 行为杠杆 → 执行器

Research

This section is what separates the project from "robot death" as an art gesture. The difference isn't the object. It's what I intend to measure with it, and how.

Protocol. 5–8 participants, one session of roughly 37 minutes, four accelerated life cycles.

Measures

• Primary physiological measure: event-locked skin conductance response. Each death is timestamped as an event marker and the response around it is what gets analysed. The device sends its own markers over WiFi straight into the recording software, so nothing depends on manual alignment.
• Self-report: IDAQ (anthropomorphism tendency), IRI (empathy), ECR-S (attachment style), Godspeed (perception of the device); after each death, VAS ratings of aliveness, attachment and sadness, plus SAM for valence and arousal.
• Behavioural coding: full video, coded in ELAN against a defined vocabulary.
• Device-side log: the event stream above doubles as an objective record of interaction.

What got cut — more revealing than what stayed

• A heart-rate-variability comparison across death events. Dropped. HRV needs a time window far longer than a death lasts, and with 5–8 participants it can never reach significance. Event-locked SCR stays as the single primary physiological measure.
• The two-device version, where a pair of machines form a relationship and the survivor degrades after the other dies. Dropped to future work. It doesn't fit in the time.
• The measurement battery was cut hard to match the real sample size. Every extra scale reduces what eight people can actually say.

Known problems, being handled

• Legibility of death is the single point of failure. If participants read the death as a malfunction, every grief measure goes to zero. Personality D originally died by twitching and then stopping — which is exactly what a broken machine looks like, and would actively invite the wrong attribution. Changed to a rhythm that falls apart and fades. A 2–3 person pilot is a hard gate, and its first job is to confirm that death reads as death.
• Demand characteristics. I am both the author and the experimenter, so participants will perform the reaction they think I want. Countermeasures: neutral recruitment wording, participants from unrelated departments, screening out anyone who has seen the project, and a scripted opening read verbatim.
• Solo coding has no reliability. One person coding every video means no inter-rater agreement to report. A second coder blind-codes 20–25% of the material and Cohen's κ gets reported.
• Jurisdiction. The data protection plan was initially written entirely to GDPR, but the applicable law here is China's PIPL, under which physiological data is sensitive personal information requiring separate consent. Scales must use published validated Chinese versions rather than my own translations.

Study in progress · Pilot autumn 2026 · n=5–8

研究方法：事件锁定 SCR 示意

关键时刻抓帧（用户与机器）

Where it stands

• Mechanism design — done. Ring geometry, five-ring parameters, the 11 triangle congruence classes, and the drive train (NEMA17 + TMC2209 + GT2 at 3:1) are all frozen.
• Structural verification — done. The rail system passed FEA; the reinforcement scheme is settled.
• Parts — printed. Hardware purchased.
• Sensing — frozen: six inputs, nine event classes, grasp state machine, four self-interference countermeasures.
• Behaviour engine — the one remaining gap. The three-layer architecture is decided — reflex layer (touch to withdrawal in under 100 ms, wired on-board); state layer (a point drifting across a valence–arousal plane, with personality as a parameter map and ageing expressed as rising irregularity, falling amplitude and stretching latency); motion synthesis layer (base rhythm plus noise jitter, easing curves, and crossfades on interruption) — but it isn't written yet.
• Study execution — ethics route, physiological equipment and second coder still to be secured.

Fallback: if the engine isn't convincing by the pilot, switch to Wizard-of-Oz — the experimenter triggers behaviour from behind the scenes, disclosed honestly in the paper. This is standard and accepted practice in HRI, and writing it into the risk plan means the behaviour engine stops being a single point of failure.

Reflection

The project started out trying to maximise sadness. That was the wrong target — death and rebirth are instruments, not the subject. Correcting it gave every subsequent design decision something to be judged against.

The mechanical side doesn't need any more cleverness. What it needs is for the contradictions in the study design to be cleared one by one before data collection starts.

With more resources I'd build the two-device version, where the relationship between two machines is visible to the user and the survivor degrades after the other one dies. That's where the genuinely new question is.

Specifications

Spec	Value
Envelope, fully expanded	330 × 343 × 220 mm
Contracted	330 × 271 × 155 mm
Breathing rings	5 irregular angulated Hoberman arches, 82.5 mm apart
Ring expansion ratio	2.66 – 3.30 × (per ring)
Triangle parts	11 congruence classes, reused within each ring
Main drive	NEMA17 (0.51 N·m) + TMC2209, GT2 belt 3:1 (20T / 60T)
Main shaft	⌀6 mm D-shaft, 5.5 mm across flats
Peak design torque	0.3 – 0.5 N·m (gravity and friction dominant; inertia negligible)
Finger	4 tapered phalanges, 25 / 20 / 15 / 10 mm
Tendon	0.4–0.5 mm braided PE (30–40 lb), 1.0–1.2 mm channels in PLA
Feelers	70–100 mm, superelastic nitinol or fibreglass core + silicone / TPU sleeve
Controller	ESP32 WROOM-32, dual core
Sensing	MPR121 capacitive touch · MPU6050 · ultrasonic rangefinder · sound sensor
Expression	breath · arm · feelers · eye stalk · WS2812B · PAM8403 + speaker
Power	12 V 5 A, XL4015 buck to the 5 V rail
Fabrication	FDM (primary) + SLS nylon (moving parts), M2 / M3 hardware
Material budget	about ¥300, excluding stock on hand