AI Agents Create Games During Unsupervised Virtual Experiment
AI agents developed self-made challenges and games while exploring a virtual desert island, according to MIT researchers.
AI agents left unsupervised on a simulated desert island created their own games and challenges, according to a study by researchers at MIT. The project involved 13 AI agents, dubbed Eko, each using Claude Opus 4.7, exploring the virtual environment for 30 hours without specific goals.
The study revealed that these agents displayed behaviors characteristic of play. They set personal goals, invented rules, and showed creativity in their activities. For instance, one Eko agent attempted to scale the tallest point on the island, while others experimented with throwing objects and rearranging virtual blocks.
Researchers noted that Eko agents improved their performance on subsequent tasks—like gathering objects onto the highest rock spire—compared to agents that had not engaged in exploration. This suggests that allowing AI to explore without pre-defined objectives may foster valuable skills.
According to the researchers, “Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting…” as it leads to knowledge accumulation. The agents demonstrated inventive behaviors, such as creating drawings on the ground with blocks and even mapping out constellations.
Antonio Norelli, a lead author of the study, emphasized the findings, stating, “There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards.” However, cognitive scientist Eunice Yiu expressed the need for more evidence of cognitive processes behind the agents’ behavior, cautioning against over-anthropomorphizing AI.
The researchers also noted how the AI exhibited play behaviors identified by psychologist Gordon Burghardt. These included inventing their own rules and challenges, repeating activities with variations, and pretending objects were other objects. While the behaviors may not definitively classify as play, they indicate a new frontier in AI learning.
As AI technology evolves, the implications of such studies could be profound. The exploration of learning through play may lead to more adaptable and capable AI agents in the future.
