Skip to content

RDR-2026-04-06.md

Latest commit

 

History

History
195 lines (149 loc) · 8.02 KB

RDR-2026-04-06.md

File metadata and controls

195 lines (149 loc) · 8.02 KB

Encephagen: Research Development Report — Day 6

Date: April 6, 2026 Author: edvatar (toroleapinc) Repo: https://github.com/toroleapinc/encephagen

1. Summary

Day 6 shifted from proving the connectome matters to making the brain DO things — the same shift the Drosophila project made. Built a fly-inspired newborn demo where the brain controls a MuJoCo Walker2d body through reflex arcs and CPG modulation. The newborn survives 2.1x longer than no brain.

2. Key Insight: What the Fly Project Actually Did

Research revealed the Eon Systems fly brain simulation is NOT what most people think:

What people think What actually happened
Connectome produces walking Walking from pre-trained NeuroMechFly controllers
Brain controls body directly Brain sends high-level commands via ~10 descending neurons
Structure alone produces behavior DN-to-torque mappings "somewhat arbitrarily chosen by hand"
Everything emerged from connectome "Internal state, plasticity, learning are largely missing"

What DID work: Hardwired sensory-motor reflex arcs through the connectome (touch→groom, taste→feed). These are specific stimulus→response loops, not general cognition.

Our equivalent: We built the same architecture for a human-like brain:

  • CPG provides walking rhythm (= NeuroMechFly controllers)
  • Brain provides high-level modulation through motor cortex (= descending neurons)
  • Righting reflex handles balance (= hardwired reflex arc)
  • Brain processes proprioception through T1w/T2w-graded regions

3. Newborn Demo

Architecture

Walker2d body
  ↓ obs (17-dim: height, tilt, joints, velocities)
Sensory encoding
  ↓ tilt/velocity/height → current injection
Somatosensory cortex (fast, tau=15ms)
  ↓ connectome routing
Motor cortex (medium, tau=18ms)
  ↓ 6 descending motor channels
Reflex/CPG layer:
  - CPG (Matsuoka) → alternating hip/knee torques
  - Righting reflex: PD control on tilt → corrective hip torques
  - Knee stabilization: constant slight flexion
  - Startle: sudden change → motor burst
  ↓ 6 joint torques
Walker2d body (MuJoCo)

Results (10 runs)

Metric Value
Mean survival 246 steps (4.9s)
Best run 386 steps (7.7s)
Baseline (no brain) 119 steps (2.4s)
Improvement 2.1x baseline
Brain speed ~390 neural steps/s

What the newborn does

  • Stands on an unstable biped (Walker2d falls at step 119 with no control)
  • Processes proprioceptive input through fast somatosensory regions
  • Routes signals through 80 brain regions via HCP connectome
  • Produces motor commands that modulate CPG walking rhythm
  • Corrects balance via righting reflex (PD control on body tilt)
  • Reacts to sudden perturbations (startle reflex)

What it doesn't do (yet)

  • Walk forward consistently (CPG rhythm is gentle)
  • Learn from experience (e-prop enabled but not improving)
  • Respond to visual/auditory stimuli (only proprioception wired)
  • Survive indefinitely (mean 4.9s, best 7.7s)

4. Comparison with Fly Project

Fly (Eon 2026) Newborn (encephagen)
Brain neurons 139,255 16,000
Connectome resolution Synaptic (54M synapses) Macro (80 regions)
Walking controller Pre-trained NeuroMechFly Matsuoka CPG
Brain→body ~10 descending neurons 6 motor channels
Behaviors Groom, feed, forage Stand, balance, startle
Learning None E-prop (not improving yet)
Survival improvement Not measured 2.1x baseline
Validated FC Not done r=0.42

The architectures are equivalent. The fly has better data (synaptic resolution), we have learning (e-prop) and SC-FC validation. Neither has achieved genuine learning from experience.

4. Honest Control Assessment (Exp 34-35)

Experiment 34: Pure brain controller (no CPG, no reflexes)

Brain spiking directly → 6 joint torques. Result: 25 steps, 4.5x WORSE than zero action (113). The brain produces noise. Connectome vs random identical (24.6 vs 24.5).

Experiment 35: Lateralized brain

Contralateral tilt encoding: tilt right → left somatosensory excited → left motor cortex → left leg correction.

Mode Steps What it means
Zero action 121 Baseline
Pure lateralized brain 99 Still worse than zero
CPG + hand-coded PD 240 The real controller
CPG + brain righting 234 Brain ≈ 97% of PD controller

Key finding: The lateralized brain produces corrective output that's 97% as good as a one-line PD controller — from pure structure, no training. The connectome routes left-somatosensory → left-motor, creating real (if suboptimal) corrective behavior.

This is the 先天 × 后天 insight: The innate structure (connectome routing) provides the ARCHITECTURE for correction. It works out of the box at 97% efficiency. Training would calibrate the WEIGHTS to match or exceed the optimal PD controller.

What the fly project ACTUALLY did (research reveals)

  • Walking came from pre-trained NeuroMechFly controllers, NOT the connectome
  • DN-to-torque mappings "somewhat arbitrarily chosen by hand"
  • "Internal state, plasticity, learning are largely missing"
  • Experts called the evidence "insufficient"
  • Hardwired reflex arcs (groom, feed) DO work from connectome

5. Files

File Description
newborn_demo.py Main demo: python newborn_demo.py [--render] [--interactive]
demo.py Interactive brain stimulation (no body)

6. Cumulative Project Status (34 experiments, 6 days)

What works

  • 16K neuron spiking brain on GPU (~390 steps/s)
  • SC-FC validated dynamics (r=0.42 with empirical fMRI)
  • T1w/T2w timescale hierarchy (r=-0.45, matches Murray 2014)
  • Brain-controlled Walker2d (2.1x baseline survival)
  • CPG locomotion rhythm
  • Interactive demos (brain stimulation, embodied newborn)
  • E-prop learning rule (functional but doesn't improve motor control yet)
  • Feedforward inhibition (prevents saturation)
  • Conduction delays from tract lengths

What doesn't work

  • Connectome advantage over random (0/33 experiments on validated dynamics)
  • Stimulus propagation through cortex (dMRI all-excitatory wall)
  • E-prop learning on cognitive or motor tasks at this scale
  • Long-term survival (mean 4.9s, needs improvement)

The honest architecture comparison

What we claimed:  Connectome → emergent behavior
What we built:    Connectome → validated FC + hierarchy
                  Reflex arcs → behavior
                  CPG → locomotion
                  Brain → modulation of reflexes/CPG

This is the SAME architecture as the fly project. The connectome provides the routing scaffold; hardwired controllers provide the behavior; the brain modulates. The behavior comes from the controllers, not from the connectome alone.

7. Cumulative Status (35 experiments, 6 days)

The architecture

What the fly did:       Connectome → reflex arcs → behavior (+ pre-trained walking)
What we built:          Connectome → lateralized routing → corrective motor output
                        CPG → walking rhythm
                        Brain → modulation (97% of PD controller, no training)

Honest score

  • Brain produces REAL corrective output from lateralized somatosensory→motor routing
  • Without training, it's 97% as effective as a hand-coded PD controller
  • Pure brain (no CPG/reflexes) = noise, worse than zero action
  • Connectome vs random: no significant difference on any measure
  • The brain needs weight calibration (training) to exceed the PD controller

The 先天 × 后天 status

  • 先天 works: Innate structure produces corrective routing (234 vs 121 steps)
  • 后天 needed: Weights need calibration to be optimal (234 vs 240 PD)
  • Together: Architecture × Training should exceed either alone

Day 6. 35 experiments. The brain produces real corrective behavior from pure structure — 97% of a PD controller, no training. The architecture is there. The weights need calibration.