Neuro-INT Mamba

A Bio-inspired Mamba Architecture for Dexterous Manipulation with Intrinsic Neural Timescales (INT).

Features

• Dual-Stream INT: Parallel fast (sensory) and slow (cognitive) streams.
• Predictive Coding: Efference copy loop for error-driven learning.
• Chandelier Gating: Inhibitory control inspired by Chandelier cells.
• Spinal Reflex: Low-level feedback for immediate response.
• Real-time I/O: $O(1)$ inference step for closed-loop control.

Installation

pip install neuro-int-mamba

Usage

import torch
from neuro_int_mamba import NeuroINTMamba

input_dims = {
    'proprio': 54,
    'tactile': 100,
    'visual': 256,
    'goal': 32
}

model = NeuroINTMamba(input_dims, model_dim=512, num_layers=6)

# Real-time control loop
states = None
predictions = None

while True:
    # Get sensory data
    p, t, v, g = get_sensors() 
    
    # Step model
    motor_cmd, states, predictions = model.step(p, t, v, g, states, predictions)
    
    # Apply motor command
    apply_motor(motor_cmd)

Development

This project uses Astral tools for high-performance development:

• uv: Package management
• ruff: Linting and formatting
• ty: Type checking

uv run ruff check .
uv run ty check .
uv run pytest

Publishing to PyPI

This project is configured to be built with hatchling and published via uv.

1. Build the package:

   uv build
   

2. Publish to PyPI:

   # Use __token__ as username and your PyPI API token as password
   uv publish
   

License

MIT