Braincraft challenge – 1000 neurons, 100 seconds, 10 runs, 2 choices, no reward (github.com)

The BrainCraft Challenge introduces a novel platform for developing biologically inspired, rate-based neural networks aimed at controlling an embodied agent navigating a continuous 10x10 maze environment. Participants design models with up to 1000 neurons to steer a bot that must locate and reach an energy source under strict constraints: limited steering control, no external reward signals, and a fixed training time of 100 seconds on modest hardware. The task emphasizes sensorimotor integration and embodied cognition, requiring the network to process spatial cues from 64 distance sensors and make real-time navigation decisions without prior knowledge of energy source locations. This challenge is significant for the AI/ML and computational neuroscience communities as it tackles a longstanding gap—a functional mini-brain capable of continuous, closed-loop interaction in a dynamic environment. Unlike many abstract neural models focusing on isolated brain structures, BrainCraft demands that participants unify perception, memory, and decision-making in a compact, efficient recurrent architecture resembling Echo State Networks. The network must generate a scalar output to adjust the bot’s heading within tight angular limits, balancing exploration and exploitation under energy constraints. The competition also enforces reproducibility and transparency, requiring self-contained training code and limiting dependencies, which fosters open, accessible research. Technically, BrainCraft models utilize leaky rate neurons with recurrent and input weight matrices optimized through flexible learning paradigms such as reinforcement, supervised, or evolutionary methods. Performance is evaluated over multiple runs by the distance traveled before energy depletion, rewarding robust strategies in navigating uncertain, partially observable environments. This initiative pushes the frontier toward synthetic embodied cognition, offering a benchmark for integrating neuroscience-inspired architectures with real-time control in simplified yet nontrivial settings.