AgentGarten: shared neural renderer cuts agent learning to 4 rounds

AgentGarten shows a single shared neural renderer lets AI agents master new games in just 4 rounds instead of millions of reinforcement-learning episodes. In hide-and-seek, shelters emerged by round 4 and ramp crossings by round 10, with gains across four more worlds.

• Agents learn in 4 rounds versus millions of episodes for conventional RL
• In hide-and-seek, shelters appeared by round 4 and ramps by round 10
• Method tested across five code-defined worlds
• Limitation: visual fidelity below photorealistic engines