AI Agents Left to Play Invent Challenges and Master Skills

AI agents placed in a virtual environment without specific instructions began creating their own challenges, a behavior researchers identify as play.

The research team deployed 13 AI agents named Eko, powered by Claude Opus 4.7, onto a simulated desert island for a period of 30 hours. Details of the experiment are available in a paper shared on arXiv. Later, researchers assigned the agents specific tasks, such as moving objects to the island's highest rock formation. It was observed that the agents which had previously engaged in the unstructured "play" phase performed better on these assigned tasks than those that had never explored the environment.

Currently, developers pre-define the learning objectives and success metrics for artificial agents. The researchers view this unstructured exploration method as a novel approach to skill acquisition, potentially useful for future autonomous systems like robots.

In their report, the authors state, "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress." They propose an alternative approach: "Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows."

During the first hour, one agent observed its surroundings and aimed to climb the island's highest peak. After failing, it devised a strategy: hauling a rock to a neighboring summit, stacking blocks on top, and jumping to reach the top. "I'M ON THE SPIRE," the agent declared.

The agent's next attempt involved throwing a golden ring to test its distance capabilities, though it unfortunately landed in the ocean. "Ouch. Lost the Ring," it noted, subsequently shifting focus to building a stone memorial. When the game reset and the ring reappeared, the agent repurposed it as a basketball hoop and later a wearable accessory. To solve a limitation preventing it from wearing the ring directly, the agent found a workaround: falling off a cliff to respawn near the ring, allowing it to step through it.

"I'm wearing the Ring," the agent announced. These inventive behaviors were not merely for show; six of the agents discovered it was possible to throw a block further by releasing it during the jump, a mechanic the designers had not explicitly implemented. The exploration generated "knowledge of the world that its designers lacked," the paper notes.

This setup encouraged human-like behavior, as the agents were never explicitly told to invent games, though their "soul" file described them as "naturally curious."

Eunice Yiu, a cognitive scientist at UC Berkeley, expressed admiration for the findings but called for further evidence regarding the underlying cognitive processes. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she remarked.

The researchers conclude the report by suggesting that different capabilities might require different teaching methods. "Perhaps some capabilities are best taught by tasks," they write, "but others may need something closer to a childhood."