Core Concepts
MineLand is a multi-agent Minecraft simulator that bridges the gap between virtual agents and real-world humans by introducing limited multimodal senses and physical needs as primary drivers of agent behavior and interaction.
Abstract
The MineLand simulator offers several key features:
Large-Scale Multi-Agent Support: MineLand can support up to 48 agents simultaneously on a mainstream consumer desktop PC, a significant improvement over previous Minecraft-based simulators.
Limited Multimodal Senses: Agents in MineLand have limited visual, auditory, and environmental awareness, forcing them to actively communicate and collaborate to fulfill their needs.
Physical Needs: Agents require fundamental physical needs like food and resources, adding a time-based aspect to their daily routines and fostering dynamic interactions.
The MineLand benchmark suite provides a wide range of tasks, including programmatic tasks, creative tasks, and hybrid tasks, enabling the evaluation of emergent multi-agent capabilities. The authors also introduce the Alex agent framework, which is inspired by multitasking theory and enables agents to handle intricate coordination and scheduling.
Experiments demonstrate that the MineLand simulator, benchmark, and AI agent framework contribute to more ecological and nuanced collective behavior, with agents exhibiting improved task performance, multitasking abilities, and cooperation when compared to agents without these features.
Stats
"MineLand is capable of supporting up to 48 agents simultaneously while providing a visual display. When visual display is disabled, the number of concurrently running agents increases to 48."
"When MineLand and Malmo both run 8 agents, MineLand's CPU and memory usage are approximately 1/3 that of Malmo's (specifically, 35.6% and 38.0%, respectively)."
Quotes
"Conventional multi-agent simulators are often under the assumption of perfect information and limitless capabilities. These idealized worlds diverge sharply from the messy reality of human interaction. This gap between simulated agents and real-world humans hinders the ecological validity and richness of social interaction within these platforms."
"By incorporating these three features, our simulator fosters the emergence of dynamic and ecologically valid multi-agent interactions."