CARL-Ω: The Complete Embodied Biological Brain 🧠🤖

The Most Intelligent Cognitive Architecture Built on a Virtual Primate Scout

---

🌟 Overview

CARL-Ω (Carl-Omega) is a biologically-inspired, multi-tiered cognitive architecture implemented on a 28-DOF physical Primate Scout robot simulated in MuJoCo.

Unlike standard AI models which are disembodied (such as LLMs), Carl operates in a closed-loop environment where he must manage physical variables (gravity, friction, collisions), negotiate homeostatic and allostatic drives (hunger, fatigue), map his surroundings, and control his 16-DOF arms to perform complex tasks like reach-grasp-and-carry.

---

🧠 The 8-Layer Cognitive Stack

Carl's brain is structured as an integrated biological hierarchy, ranging from raw physics up to conceptual abstraction:

┌─────────────────────────────────────────────────────────────────┐
│  LAYER 7: CONCEPT GENESIS (Phase 16)                            │
│  Abstractions computed from raw surprise recurrence patterns    │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 6: THE IMAGINATION (Phase 15)                            │
│  Hallucination Engine: Dreamer world model offline simulation   │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 5: CAUSAL REASONING (Phase 14C)                          │
│  Counterfactual Question Generator + Causal Adjacency Scaffold │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 4: THE WITNESS (Phase 14B)                               │
│  Metacognitive memory: monitors own failures and penalizes errors │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 3: UNCERTAINTY MAP (Phase 14A)                           │
│  2D Spatial Navigator (Danger × Uncertainty matrix)              │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 2: THE BIOLOGICAL BRAIN (Phase 13)                       │
│  Neuromodulators (DA/NE/5-HT) │ Hebbian Plasticity │ Sleep      │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 1: TWO-SPEED MEMORY (Phase 11)                           │
│  Working Memory (Fast) ⊗ Long-Term Memory (Slow)                │
├─────────────────────────────────────────────────────────────────┤
│  LAYER 0: PHYSICAL REALITY                                      │
│  MuJoCo Physics, 28-DOF body, motor lag, and sensor noise       │
└─────────────────────────────────────────────────────────────────┘

---

🧬 Biological Engine Breakdown

1. The Endocrine System & Plasticity

• Neuromodulators: Carl's learning rates and arousal thresholds are dynamically guided by internal transmitters:
  • Dopamine (DA): Signals reward prediction error and gates learning.
  • Norepinephrine (NE): Signals surprise and increases arousal (thermodynamic action noise).
  • Serotonin (5-HT): Acts as a safety guard to stabilize memory.
• Synaptic Plasticity: Utilizes Reward-Modulated Spike-Timing-Dependent Plasticity (R-STDP). Associations strengthen when pre-synaptic neurons fire right before post-synaptic ones (LTP) and weaken when reversed (LTD).

2. Holographic Sub-Cortical Memory (HDC)

• Maps observations to a 40,000-dimensional bipolar hypervector space ({-1, +1}^D) split into independent sub-cortical registers:
  • M_spatial — environmental traces
  • M_affective — emotional signatures
  • M_procedural — motor actions
• Memories are superposed using bundling (+) and binding (⊗) algebraic operations without crosstalk, pruning noise during offline sleep consolidation.

3. Hippocampal GPS Navigation

• Simulates Nobel-Prize winning Grid & Place cell biology (O'Keefe & Moser 2014):
  • Grid Cells: 3 modules of hexagonal tiling grid cell populations.
  • Place Cells: 200 place cells mapping local coordinates.
  • Coordinates path integration (dead reckoning) and maps safety gradients.

4. Liquid Time-Constant Motor Cortex (LTC)

• Driven by a recurrent LTC neural network (imitation-learned via Behavioral Cloning and fine-tuned with PPO):
  • Learns dynamic, physical reaching and grasping profiles.
  • Adjusts its internal time constants dynamically according to environmental variables.

---

📁 Repository Structure

carl-simulation/
├── brain/                    # Bio-computational components
│   ├── carl_stdp.py          # R-STDP & dopamine-modulated learning
│   ├── carl_grid_cells.py    # Hippocampal Grid & Place cells
│   ├── carl_physarum.py      # Slime-mold Steiner path optimizer
│   └── carl_omega_extensions.py
├── GENESIS/                  # Core simulation environment
│   ├── carl_scout/           # Primate Scout agents and controllers
│   │   ├── carl_cortex.py    # Main sensorimotor integration backbone
│   │   ├── carl_agent.py     # Actor-critic brain, Drives & HDC memory
│   │   └── carl_scout_ppo.py # PPO Arm-reaching fine-tuning
│   └── carl_autonomous.py    # Closed-loop reach-carry-place state machine
├── docs/                     # Research roadmaps, architecture, and manifests
├── dashboard/                # Live performance visualizations
└── tests/                    # Kinematics and physics stability diagnostics

---

🛠️ Getting Started

1. Setup Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Run the Autonomous Demo

Execute the full pick-and-place sequence using the cognitive cortex:

python GENESIS/carl_autonomous.py

This runs the closed-loop task sequence: $$\text{SEARCH} \to \text{APPROACH} \to \text{REACH} \to \text{GRASP} \to \text{LIFT} \to \text{CARRY} \to \text{PLACE} \to \text{SEARCH}$$