biological-alignment-benchmarks/biological-alignment-gridagents-benchmarks

Implements eleven gridworld benchmarks across three developmental stages—basic biologically-inspired dynamics, multi-objective learning, and multi-agent cooperation—addressing overlooked safety aspects like homeostasis, diminishing returns, and resource sharing. Built on the Extended Gridworlds framework, it provides training implementations for Stable Baselines 3 RL agents and LLM-based agents, with extensible templates for custom agent development. Designed for empirical evaluation of how agents learn utility functions aligned with biological and economic principles rather than pure reward maximization.