Date: April 7, 2026 Author: edvatar (toroleapinc) Repo: https://github.com/toroleapinc/encephagen Duration: 7 days, 37 experiments
A miniature human newborn simulation: 16,080 spiking neurons controlling a 3D Humanoid body through biologically correct subcortical circuits. Zero learning. Pure innate behavior.
CORTEX (16,000 LIF neurons, 80 brain regions)
- HCP structural connectivity (neurolib80, 6220 continuous weights)
- T1w/T2w timescale gradient (sensory 10ms → frontal 30ms)
- Conduction delays from tract lengths (7-238ms)
- SC-FC validated (r=0.42 against real fMRI)
- Status: OBSERVER — not controlling the body at birth
BRAINSTEM (hardwired reflex arcs)
- Righting reflex (tilt → corrective torque)
- Moro reflex (sudden drop → arms extend then flex, 2 phases)
- Startle reflex (sudden change → whole-body jerk)
- Palmar grasp (constant elbow flexion)
- ATNR/fencing (tilt → asymmetric arm posture)
- Breathing CPG (~40 breaths/min, abdomen oscillation)
- General movements (Prechtl's — multi-frequency fidgeting)
- Stepping drive (upright → activate spinal CPG)
- Basal ganglia gating (reflex prioritization)
SPINAL CPG (80 LIF neurons, identified interneuron classes)
- Shox2 rhythm generators (10 neurons, pacemaker)
- V1 inhibitory (8 neurons, flex-ext alternation)
- V2a excitatory (6 neurons, drives commissural)
- V2b inhibitory (8 neurons, flex-ext alternation)
- V0d inhibitory (6 neurons, left-right alternation)
- V0v excitatory (6 neurons, left-right alternation)
- V3 excitatory (6 neurons, symmetry)
- Motor neurons (20 neurons, flex/ext per side)
- Architecture: Rybak et al. 2015, Danner et al. 2017, Kiehn 2016
- Weights: CMA-ES optimized (40 generations, fitness 5.5→2.4 verified)
BODY (MuJoCo Humanoid-v5)
- 17 actuated joints (torso, 2 arms, 2 legs)
- Survival: 1.2x baseline (Humanoid) / 2.2x baseline (Walker2d)
| Behavior | Neural Basis | Status |
|---|---|---|
| Righting reflex | Brainstem vestibular → reticulospinal | ✅ Working |
| Stepping reflex | Spinal CPG (80 spiking neurons, V0/V1/V2a/V2b/V3) | ✅ Working |
| Moro reflex | Brainstem → bilateral arm extension then flexion | ✅ Working |
| Startle reflex | Brainstem reticular → whole body | ✅ Working |
| Palmar grasp | Spinal C5-T1 → constant elbow flexion | ✅ Working |
| ATNR/fencing | Brainstem → asymmetric arm posture | ✅ Working |
| Breathing | Brainstem medullary CPG → abdomen rhythm | ✅ Working |
| General movements | Brainstem → multi-frequency whole-body fidgeting | ✅ Working |
| Knee stabilization | Spinal → constant slight flexion | ✅ Working |
| Withdrawal | Spinal → flex stimulated limb | |
| Babinski, plantar grasp | Spinal → toe movements | ❌ No toes on body |
| Rooting, sucking | Brainstem trigeminal → mouth | ❌ No mouth on body |
| Swimming | Brainstem/spinal → paddling | ❌ No water environment |
Missing behaviors are due to missing body parts (toes, mouth) and environments (water), NOT missing neural circuits.
- SC-FC validated dynamics (r=0.42 vs real fMRI) ✅
- Timescale hierarchy matching Murray 2014 (r=-0.45) ✅
- Lateralized corrective routing (97% of PD controller without training) ✅
- Cognitive advantage over random wiring (0/33 experiments on validated dynamics)
- Stimulus propagation through cortex (blocked by all-excitatory long-range)
- Learning advantage (neither connectome nor random learns at this scale)
- Sustained gait from 80-neuron spiking CPG with identified interneuron classes ✅
- This is the fly/worm approach: correct architecture + calibrated weights → behavior
Behavior comes from specific circuits, not from general connectivity. The macro-scale connectome (80 cortical regions) provides a validated scaffold. Specific subcortical circuits (80 spinal neurons, brainstem reflex arcs) produce the actual behavior. This is biologically correct: newborn behavior is subcortical.
| C. elegans (BAAIWorm) | Fruit Fly (Eon 2026) | Encephagen | |
|---|---|---|---|
| Brain neurons | 302 (all identified) | 139,255 (all identified) | 16,080 (80 cortical populations + 80 spinal identified) |
| Connectome resolution | Synaptic (EM) | Synaptic (EM) | Macro (dMRI) + spinal identified |
| Body | 3D worm (Sibernetic) | NeuroMechFly (MuJoCo) | Humanoid-v5 (MuJoCo) |
| Walking source | Connectome + parameter optimization | Pre-trained NeuroMechFly controllers | CMA-ES optimized spiking CPG |
| Innate behaviors | Chemotaxis, locomotion | Grooming, feeding, foraging | 10 neonatal reflexes |
| Learning | None | None | None (innate only) |
| SC-FC validation | Not applicable | Not done | r=0.42 ✅ |
| Weight calibration | Parameter optimization | EM synapse counts (evolution) | CMA-ES (spinal) + Brunel (cortical) |
- SC-FC validated cortex matching real human fMRI
- T1w/T2w timescale gradient from HCP data
- Multiple distinct reflexes (10 behaviors, not just locomotion)
- Cortex as observer with real human connectome (ready for developmental learning)
- Synaptic-resolution connectome (each neuron identified)
- Behavior emerging purely from connectome (no hand-coded reflexes)
- Neurotransmitter identity per synapse
We have reached the limit of what pure innate structure can produce:
Done (先天):
- All reflexes a real newborn has (limited by body parts, not neural circuits)
- Breathing rhythm, general movements, stepping
- Cortex processing sensory input through validated human connectome
- 16,080 spiking neurons across 3 hierarchical levels
Requires 后天 (learning):
- Better balance (real newborns improve over days/weeks)
- Cortex taking over from brainstem (corticospinal myelination, 2-4 months)
- Primitive reflexes disappearing (cortex learns to inhibit them)
- Voluntary movement replacing reflexive movement
- Any skill beyond hardwired reflexes
The fly project also stopped at innate behavior — no learning. The difference: their innate behavior comes from the connectome; ours comes from hand-coded brainstem + CMA-ES optimized spinal CPG. Both are biologically correct for their scale — the fly has synaptic resolution, we have macro-scale + specific subcortical circuits.
# Full newborn (Humanoid, all reflexes, spiking CPG, video)
python newborn_full.py --video
# Walker2d version (simpler body, longer survival)
python newborn.py --spiking-cpg --video
# Interactive brain stimulation (no body)
python demo.py
# Generate comparison video
python visualize_humanoid.py| # | Experiment | Key Finding |
|---|---|---|
| 1-4 | Wilson-Cowan phase 1 | Degree→hierarchy, wiring→FC |
| 5 | Spiking hierarchy | Matches Murray 2014 |
| 6 | STDP habituation | Repetition suppression |
| 7-9 | Embodied learning | Invalidated (motor death) |
| 10 | Pendulum learning | 5 approaches failed |
| 11 | Spontaneous body | Rhythmic twitching emerges |
| 12-14 | CPG + body | 0.98 Hz gait, crawling worm |
| 15-18 | Cognitive functions | Conditioning, discrimination, memory, integrated |
| 19-20 | Walker2d control | Brain helps 78% |
| 21-24 | Connectome vs random | Structure helps organization, not cognition |
| 25-26 | SC-FC validation + tuning | r=0.388 (tvb96), r=0.42 (neurolib80) |
| 27 | Validated connectome vs random (tvb96) | 1/5 structure, 3/5 random wins |
| 28 | tvb66 tuning | Continuous weights, log-transform |
| 29 | Validated (neurolib80) | 0/4 significant — no structural advantage |
| 30 | Innate dynamics | Stimulus doesn't propagate (dMRI wall) |
| 31 | Learning scaffold | Neither brain learns |
| 32 | Newborn closed-loop | Body too stable |
| 33 | Walker2d brain control | 210 vs 119 baseline, no structural advantage |
| 34 | Pure brain controller | Brain = noise (25 steps, 4.5x worse than zero) |
| 35 | Lateralized brain | 97% of PD controller from pure structure |
| 36 | ES spiking CPG optimization | Intermittent alternation |
| 37 | CMA-ES spiking CPG | Sustained oscillation (fitness 5.5, verified 2.4) |
7 days. 37 experiments. 16,080 spiking neurons. 10 innate behaviors. The 先天 phase is complete. The newborn is alive.