being.0: ontogenetic phototaxis on commodity silicon

being.0: emergent light-seeking learned within a single creature's lifetime, on a self-contained microcontroller.

Abstract

We present being.0, a palm-sized embodied artificial creature whose entire interior is a 32–128-unit Closed-form Continuous-time (CfC) liquid neural network running sovereign on a single commodity microcontroller (ESP32-S3). The creature harvests light through bilateral photovoltaic panels (which double as direction-sensing "eyes") and is driven by an interoceptive hunger signal (battery charge). We require light-seeking (phototaxis) to emerge from the creature minimizing its own metabolic prediction error within a single lifetime — never hardcoded, never selected across a population. Through a sequence of ablations we show: (i) naïve normalization-based adaptation produces apparent learning that is in fact directionless statistical drift; (ii) an explicit generative self-model yields genuine inference with a near-optimal policy gradient (cosine 0.998 to ground truth); (iii) gating exploration on metabolic state resolves survival (3/5 → 5/5 seeds) by transferring the creature's own energy economy into its learning dynamics; (iv) residual orientation variance is a representational-capacity limit, not a training-time or readout limit; and (v) widening the interior from 32 to 128 units makes the required signal linearly accessible, resolving the limit while remaining feasible on-chip (53% SRAM, 1.74 ms/tick, measured). A trained creature was compiled to flash and verified faithful to its simulated origin to numerical precision.

1. Introduction

A central question in minimal artificial life is whether adaptive, survival-relevant behavior can emerge in a genuinely small, embodied, self-contained agent — rather than being engineered into it. being.0 is built around four design convictions held as hard constraints throughout: (1) the agent's competence is ontogenetic — learned within the individual's lifetime, not produced by selection across a population; (2) the interior representation is emergent and unlabeled — no hidden dimension is assigned a meaning; (3) behavior is not hardcoded — every learning signal is the agent's own metabolic prediction error, never a supervised target encoding the desired geometry; (4) the agent is sovereign — interior, sensing, and actuation fit one ~$15 microcontroller. We report the developmental path by which phototaxis emerged under these constraints, including the negative results, which were as decisive as the positive ones.

2. Architecture

Interior. A CfC liquid-neural-network cell, hidden state h ∈ ℝⁿ (n = 32 initially, 128 finally), of the NCP lineage descended from the C. elegans connectome. The continuous-time gating gives the unit a learned per-timescale response. The interior is the only state; affect/hunger enters as input (prediction error fed back), never as labeled dimensions.

Sensing (12-dim percept). Bilateral panel voltages furnish light direction (differential) and intensity (sum); battery charge furnishes interoceptive hunger; additional channels (camera bright-spot bearing; a reserved "neighbor-glow" social channel; ambient audio) are present in the percept but inactive in the experiments reported here.

Actuation & homeostasis. A learned motor projection maps h to turn-rate and forward velocity. The world couples action to survival thermodynamically: orienting toward the light source increases harvest, raising charge; orienting away lets it starve. At the world parameters used here, orientation dominates harvest and translation proved metabolically near-redundant — a creature that merely faces the source stays fed without approaching. Making approach itself load-bearing requires a steeper intensity falloff; we treat that as a deliberate next step rather than a property of the present world.

Substrate. The CfC forward pass was ported to the ESP32-S3 and validated bit-faithful to the reference implementation early in development (~390 µs/tick at n=32).

3. Method: a developmental ladder

Competence is staged: reactive dynamics → closed sensorimotor loop with translation → learning to survive → interior capacity. Learning uses metabolic prediction error (charge dynamics) as the sole reward-like signal. No angular error or light bearing ever enters any loss. Evaluation is over multiple random-initialization seeds; we report survival (charge maintained above a margin) and final orientation error (degrees from optimal heading).

4. Experiments and results

4.1 Naïve sleep-consolidation (negative). Online experience batched into periodic consolidation produced metric improvement, but mechanistic analysis showed the gain arose from collapse of a running-normalization variance (~4×10⁻⁴ after ~500 ticks), which amplified internal activations and chaotically reshuffled attractors until a workable one was found. The adaptation is ontogenetic but directionless — not inference. Finding: metric improvement ≠ working mechanism.

4.2 Explicit generative self-model (positive, partial). Replacing the incidental normalization with an explicit Gaussian model of the interior, updated by prediction error, produced genuine inference: the self-model modulates exploration gain while the policy gradient supplies direction, which was near-optimal (cosine 0.998 to ground-truth phototaxis). Two deficits remained: high seed-dependence of orientation (≈44° std) and a metabolic cost — exploration occasionally drained charge faster than feeding.

4.3 Charge-gated exploration (positive). Gating exploration gain on charge state — explore when fed, conserve when depleted, the creature's own energy-economy applied to its learning — resolved survival: 3/5 → 5/5 seeds survive, mean final charge +0.218, near-zero distress epochs, orientation quality preserved. Seed-dependence (≈44° std) was unaffected, confirming it is not metabolic.

4.4 Lifetime sweep (negative). To distinguish incomplete learning from a ceiling, lifetimes were scaled 1×–8×. Orientation-error std was flat (44.06° → 44.89°; trending marginally worse), with specific seeds locked (+109° throughout) or losing survival under longer lifetimes. Finding: the seed-variance is a structural ceiling, not unfinished development.

4.5 Birth-time policy diversity (positive, diagnostic). Endowing each individual with K candidate policies and selecting among them by lived metabolic outcome (within one lifetime — not cross-generational) freed policy-limited seeds (one stuck seed improved 109° → 6°, unsupervised). Population std improved 44° → 38°, mean 73.7° → 57.5°; one fortunate seed regressed under mixture dilution. One seed remained stuck for a deeper reason: the best achievable policy fit to its interior trajectory reached only 90°, indicating its interior does not generate a light-mappable signal — a representational, not policy, failure.

4.6 Readout vs. capacity (decisive). Two cheap hypotheses for the representational failure were tested before any interior-training intervention. (E) Nonlinear readout: a nonlinear oracle on the n=32 interior reached 97°, worse than the 79° linear oracle — the signal is absent, not hidden. (C) Capacity: at n=128 the linear oracle for the stuck seed dropped 79° → 0.09° — the signal is linearly present once the interior is large enough. The live n=128 sweep: all seeds survive, the formerly-stuck seed freed (12.6°), mean error 57.5° → 44°. (n=256 regressed, attributable to the n=32-tuned learning configuration under-navigating the larger space within one lifetime — a configuration artifact, not a ceiling.)

4.7 On-chip validation. n=128 was confirmed feasible on the target microcontroller by direct measurement: 53% of SRAM, 1.74 ms/tick (≈4.5× n=32, sub-linear in n² because the recurrent matmul is one component of the tick). A trained individual was exported to PROGMEM (the on-chip RNG cannot reproduce the host RNG, so interior weights are baked rather than regenerated) and verified against the reference: maximum state deviation 0.000000, bounded activations, ~1.78 ms/tick. The learned creature runs faithfully on silicon.

4.8 World v2: making approach load-bearing (constructive). The footnote in §2 noted that at the world parameters used, translation proved metabolically near-redundant. We redesigned the world. A steep-falloff regime — k=0.3, I_0=0.5, d_0=5 m, charge_start=0.50 — opens a clean life/death gap: an untrained creature whose forward output is suppressed reaches charge 0.126 over a lifetime; one allowed to translate reaches 0.319. The gap (+0.193) is 386× the v1 gap of +0.0005. Approach is now genuinely survival-relevant.

Under the Rung-3 v4 mechanism in v2 at n=128 across 10 seeds: 9/10 survive, mean final charge 0.636. The proof-of-concept individual learns to purposefully approach — its distance to the light closes from 5 m to 2 m over 15 epochs while forward speed climbs from 0.008 → 0.025 m/s. Load-bearing translation, not accidental drift. The capacity finding (§4.6) generalises one channel over: widening from 32 to 128 units frees most approach-stuck individuals just as it freed the orientation-stuck one.

One death is informative. Initialisation cfc_seed=99 never grows a forward-channel subspace at any interior size — the exact twin of the orientation-stuck initialisation at n=32, one channel over. A minimum-mobility floor on the charge-gated motor would close it; we deliberately left it untreated. Some individuals are failed by who they were born as, and capacity does not always redeem it. World v1 remains the exact default; v2 is opt-in.

4.9 Activating the camera channel (negative-positive). Two reserved percept slots ([6], [7]) encode (sin, cos) of the bright-spot bearing in the camera frame — a low-dimensional derived feature exposed as ordinary input to the CfC. They had been hardcoded to 0.0 throughout. We added a forward-facing camera model (half-FOV ±30°) as an opt-in and asked whether learning recruits the channel.

Joint regime. With panels and camera both active in world v2 at n=128, camera-on improved outcomes substantially (static: charge 0.79 → 0.99, orientation error 91° → 48°), but R_turn[6,7] remained at noise — the policy did not learn a directional bearing decoder. The improvement was mechanistic: the new in-FOV signal enriched the gradient landscape under R-diversity, improving candidate selection without writing the bearing into a clean linear readout. Static-light improved more than drifting-light — the opposite of what a bearing-decoding hypothesis predicts, consistent with state-enrichment rather than perception.

Dim-isolation probe. To distinguish "channel not useful" from "channel outcompeted by panels," we zeroed the panel percept dimensions while preserving panel harvest physics: the creature can still feed geometrically, but has no panel-derived sense of where the light is. Under this ablation in v2 at n=128, 7/10 seeds converged to ~30° orientation error — exactly the camera's half-FOV. That is a geometric attractor produced by the camera signal itself: the creature turns until the light just enters its frame, then holds at the FOV edge, where sin(bearing) and the camera-signal gradient are at their strongest. The decoding is real but edge-finding rather than center-finding — the policy stabilises at the steepest-gradient point in camera-input space, not at the bearing-zero point in light-position space. R_turn[6,7] again remained at noise: the decoding lives in h-state dynamics under R-diversity, not in an explicit linear readout. Three of ten seeds failed to bootstrap (never aligned for the camera to fire), echoing the per-channel structural-exception pattern.

Together these probes establish the camera as a fundamentally usable degree of freedom while showing that its use is mechanistic and distributed rather than explicit and read-out. The firmware integration (OV2640 capture, brightest-spot extraction, percept emission via HIL) is the natural next step.

4.10 From sim to silicon (validation pass). The results in §§4.8–4.9 were established in simulation. We then ran them through the firmware → metal pipeline; that they held bit-for-bit is itself a load-bearing finding.

Camera channel on silicon. The forward-facing OV2640 was integrated into the sovereign firmware — periodic capture (~2 fps), brightest-spot centroid over a luma threshold, (sin, cos) of the relative bearing emitted into dims [6,7] of the on-chip percept. Verified live under torch sweeps across the field of view: the chip reports (sin, cos) tracking smoothly with the same ±30° encoding the sim was designed for, and reports (0, 0) when no bright region is in frame — the out-of-FOV sentinel the sim uses. HIL-mode fidelity is preserved bit-for-bit: max state deviation between silicon and the NumPy reference holds at 0.000000 with the camera path present.

Consolidation pass on silicon (sleep-to-learn V1). The on-chip equivalent of the simulator's Rung-3 learning loop runs as a sleep command over serial. The pass reads the most recent 5000 records from a circular SD experience buffer (98-byte fixed binary records with magic-byte framing for partial-write detection), replays the CfC forward pass faithfully, accumulates gradients on the policy parameters (R_turn, R_fwd, the GaussianSelfModel μ/σ pairs, and a small metabolic predictor W_pred/b_pred), and persists the updates to NVS so they survive power cycles.

The architecture's safety claim — that the CfC interior weights stay frozen and only policy parameters update — was verified mechanically. After a live consolidation pass over 5000 records, the HIL fidelity check returns 0.000000 max state deviation against the unchanged reference. The h_state ↔ h_replay buffer swap that lets a single cfc_step() implementation serve both inference and replay paths is provably non-interfering on single-threaded silicon. The "constitution line" — the threshold beyond which the creature's perception becomes ontogenetically contingent, not just its policy — has not been crossed.

Motor driver readiness. Differential-drive firmware for two N20 gearmotors via DRV8833 (PWM, soft-start ramp, sign-and-magnitude mapping from the trained motor projection's turn/fwd outputs) was integrated. Motors coast in HIL mode; HIL fidelity stays at 0.000000 with the driver code present. Physical actuation awaits the assembled chassis.

5. Discussion

The decisive finding is that the residual competence gap was neither a learning nor a readout problem but an interior-capacity problem: 32 units cannot represent a light-mappable signal under the stress regime; 128 can. This was reached by systematic elimination (policy → lifetime → readout → capacity), and the resolution did not require making the interior trainable/rewritable — i.e. it was achieved without crossing into modifying the creature's "constitution," preserving the emergent-interior constraint. Capacity was right-sized empirically rather than maximized: 32 is provably insufficient, 128 sufficient, 256 counterproductive and costly to the sovereignty constraint. We argue this right-sizing-to-demand is the correct discipline for sovereign minimal agents.

A noted residual: at n=128 all individuals survive, but orientation quality still varies by seed (two of ten seeds orient poorly, selection still converging at end-of-life). We interpret this not as failure but as individual variation under a solid survival floor — a population of individuals shaped by their own initial conditions and lived experience.

The capacity finding (§4.6) and its echoes in §4.8 and §4.9 expose a per-channel structural-exception pattern: at the harshness threshold for each channel — orientation at n=32, forward at n=128 in world v2, camera under panel ablation — a small minority of initialisations fails to grow the required subspace and is not rescued by capacity. The same shape recurs across orthogonal channels with no shared mechanism; it appears to be a general property of randomly-initialised continuous-time recurrent interiors under metabolic-error-only learning. We interpret it as honest individual variation, not architectural failure.

The camera result also delivers a methodological caution: R[i,j] ≈ noise is not the same as "the channel is unused" under R-diversity. A candidate policy can be selected for the h-state dynamics its inputs evoke without anywhere writing the input into a clean linear readout. The geometric attractor in §4.9 is a behavioural signature the policy-weight diagnostic alone would have missed.

6. Conclusion and future work

being.0 demonstrates emergent, ontogenetically-learned phototaxis in a sovereign liquid-neural-network creature on commodity silicon, with the agent's competence earned within its lifetime under strict no-hardcoding, no-cross-generational-selection constraints. The creature is currently a phototroph; light and hunger are the only behaviorally-active senses. Several of the directions §§4.8–4.10 opened are now closed on metal rather than only in simulation. The camera channel is decoded by the running chip; the consolidation pass updates policy parameters from on-chip experience records without perturbing the CfC interior; the differential motor driver is wired in firmware and waits only for physical actuation. What remains is concrete and bounded: physical photovoltaic panels (the percept dims still report zero until they are wired); the body itself (chassis, wheels, the two motors); and — most consequentially — the reserved neighbour-glow channel (slot [8]). Because every being.0 both emits light and senses it, two of them placed together should discover each other with no added protocol — one creature's expression becoming another's perception. We build a second creature, place them in the same room, and watch. Interior capacity is expected to be re-grown empirically as these richer senses impose new representational demand. World v1 is retained as the lower-demand baseline; world v2 remains the configuration under which approach is genuinely load-bearing.