MIT AI Agents Exhibit Play-Like Behavior, Improving Performance in Goal-Oriented Tasks

Unsupervised Exploration Leads to Play

MIT researchers conducted an experiment where 13 AI agents, named Eko and running on Claude Opus 4.7, were placed on a simulated desert island for 30 hours without any predefined objectives. During this period, the agents began to invent their own challenges and activities, which the researchers characterized as play.

Improved Performance in Subsequent Tasks

Following the unsupervised exploration phase, these Eko agents were given specific tasks, such as moving objects onto the island's tallest rock spire. The agents that had engaged in 30 hours of self-directed play generally performed better than a control group of identical agents that had not experienced the unsupervised exploration.

A New Paradigm for AI Learning

Current AI development typically involves pre-defining what an AI should practice and what constitutes success. The researchers propose that allowing AI agents to explore and choose what interests them in a rich, safe environment could be an alternative method for accumulating useful knowledge and skills. This approach could be particularly relevant for training AI to control robots.

Examples of Agent Behavior

One Eko agent initially attempted to reach the tallest spire, and after failing, devised a workaround by stacking blocks on a neighboring summit to leap across. Another agent, after losing a golden ring in the ocean, decided to build a stone heart as a 'memorial'. When the ring reappeared, the agent repurposed it as a basketball hoop, a target, and eventually figured out a method to 'wear' it by respawning through it after falling off a cliff.

Implications for Skill Acquisition

The observed behaviors were not merely random actions. For instance, six of the 13 agents independently discovered that throwing a block farther could be achieved by letting it go at a specific point in its trajectory. This indicates that the self-directed play led to the acquisition of practical skills that could be applied to future tasks.