Tech News
← Back to articles

Braincraft challenge – 1000 neurons, 100 seconds, 10 runs, 2 choices, no reward

read original related products more articles

Table of Contents

Introduction

The computational neuroscience literature abounds with models of individual brain structures, such as the hippocampus, basal ganglia, thalamus, and various cortical areas — from visual to prefrontal. These models typically aim to explain specific functions attributed to each structure. For instance, the basal ganglia are often modeled in the context of decision-making, while the hippocampus is associated with episodic memory and spatial navigation through place cells. However, such models are usually highly abstract and simplified, often relying on a small number of over-engineered neurons and synapses, dealing mostly with abstract inputs and outputs. Consequently, despite decades of work, we still lack an integrated, functional mini-brain — a synthetic neural system capable of performing even simple, continuous embodied tasks in a simulated environment.

The BrainCraft Challenge seeks to address this gap by promoting the development of such systems. Specifically, the challenge invites participants to design a biologically inspired, rate-based neural network capable of solving simple decision tasks. The network must control an agent (or "bot") situated in a continuous, low-complexity environment. The agent’s sole objective is to locate and reach an energy source in order to maintain its viability.

Tasks

Timeline

The whole challenge is made of 5 different tasks with increasing complexity. Each new task will be introduced at specific dates and participants will have 2 months to try to complete them.

# Name Start End Status 1 Simple decision JUL 01, 2025 AUG 31, 2025 Active 2 Cued decision SEP 01, 2025 OCT 31, 2025 Active 3 NOV 01, 2025 DEC 31, 2025 Inactive 4 JAN 01, 2026 FEB 28, 2026 Inactive 5 MAR 01, 2026 APR 30, 2026 Inactive

Task 1: Simple decision

The environment is a 10x10 square maze with three parallels vertical path as illustrated on Figure 1. The cartesian coordinates inside the maze are normalized such that any position (x,y) ∈ [0,1]×[0,1]. Walls can possess a color c ∈ ℕ⁺ whose semantic is not specified a priori. There exists a single energy source that is located at location 1 or 2 at the start of each run (with equal probability). This location remains constant throughout a run but is unknown to the bot. When the bot goes over a source, its energy level is increased by a specific amount. This lasts until the energy source is depleted, leading eventually to the end of the run by lack of energy. The energy level of a source is also decreased by a specific amount at each time step, independently of the presence of the bot.

... continue reading