The agents began experimenting with their surroundings, attempting difficult climbs, arranging objects and inventing challenges. Researchers identified several behaviors associated with play, including repeated experimentation, self-directed goals and attempts to overcome obstacles.
One agent became determined to reach the island's highest rock formation. After unsuccessful attempts, it devised a different approach, moving a rock onto a nearby summit and stacking additional blocks to create a launching point. The improvised structure allowed it to reach the previously inaccessible location.
Another sequence involved a golden ring that an agent accidentally threw into the ocean. Unable to retrieve it, the agent constructed a heart-shaped arrangement of stones as a memorial. When the simulation reset and restored the ring, the agent found new uses for it, including as a target and a makeshift basketball hoop.
The agent eventually attempted to position the ring around its virtual body. When the environment prevented that action directly, it discovered that jumping from a cliff would cause it to reappear in a location that allowed it to pass through the ring.
Beyond these unusual activities, the agents uncovered techniques that the researchers had not anticipated. Six of the 13 independently discovered that releasing a block during a jump increased the distance it could be thrown. The behavior emerged through experimentation rather than an explicitly assigned challenge.
The researchers viewed these discoveries as evidence that unrestricted exploration can reveal useful properties of an environment, including interactions that its developers may not have considered.
To assess whether the experience produced measurable benefits, the team subsequently assigned the agents structured challenges. One required moving as many objects as possible onto the island's tallest rock formation.
Agents that had previously explored the island generally outperformed otherwise identical agents without that experience. Their earlier interactions appeared to provide knowledge they could apply when completing the new assignments.
The results suggest that AI agents may acquire reusable problem-solving techniques through exploration, rather than relying exclusively on activities selected in advance by developers.
Most AI development methods involve establishing an objective and evaluating a system against predetermined measures of success. The MIT researchers examined whether allowing agents to choose their own activities could help them discover capabilities that structured exercises might overlook.
The team sees potential applications in robotics, where machines may encounter unfamiliar objects, changing environments and problems that developers cannot fully anticipate. Allowing AI systems to experiment in simulations could help them build practical knowledge before attempting more demanding tasks.
However, the study does not establish that the agents experienced play in the same way humans do. Their behavior was influenced by their initial instructions, including a configuration file describing Eko as “naturally curious.” Although the agents were not directed to invent games, that characterization may have encouraged exploratory behavior.
Eunice Yiu, a cognitive scientist at the University of California, Berkeley, also questioned whether the observations were sufficient to establish genuine play.
“It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Yiu told MIT Technology Review.
The distinction matters because behavior resembling human curiosity does not necessarily demonstrate comparable internal processes. An AI system can generate playful actions and language without establishing that it experiences enjoyment, boredom or other human emotions.
The findings also remain preliminary. The research was posted to arXiv before peer review, involved a relatively small group of agents and took place entirely within a simulated environment. Additional testing would be needed to determine whether similar benefits appear across other AI models, more complex environments or physical robots.
Unstructured exploration also requires time and computing resources without guaranteeing useful discoveries. Researchers would need to determine which environments and experimental conditions produce reliable improvements before applying the technique more broadly.
Nevertheless, the experiment offers a measurable result beyond the agents' unusual behavior: systems given time to explore independently generally performed better when subsequently confronted with defined objectives.
The researchers suggest that structured assignments and exploratory learning could serve complementary purposes in AI development. While some capabilities can be taught through specific tasks, others may emerge when agents are given opportunities to experiment, encounter unexpected outcomes and retain what they discover.
This analysis is based on reporting from TechCrunch & Medium.
Image courtesy of ZME Science.
This article was generated with AI assistance and reviewed for accuracy and quality.
