AgentMe autonomously observes and learns natural language tasks, designed to gamify human-in-loop multi-agent training and evaluation.

Language to executable actions generative model, Thinking-Action Control for player NPC agent interaction, Decentralised Gym-compatible game spaces.

Divergent LLMs for Language-Action Generation

Models for language and instruction generation, separated reasoning and action models for better execution rate.

Think2Action UI for Human-Agent Interaction

AgentMe enables evaluation of AI agents with real human partners, through a Think2Action (T2A) UI.

In-Game and Real World Tasks

AgentMe tasks exhibit characteristics of everyday tasks, such as spatial, temporal, and heterogeneous agent capability constraints.

Challenging and Rewards to Improve Agent Capabilities

AgentMe gamifies reinforcement learning as play to annotate that makes RL with Human Feedback sustainable and educational.