MIT researchers say AI agents play on virtual island after 30 hours

The authors describe the experiment in an arXiv paper and argue that allowing agents to explore a rich, safe world without predefined objectives can produce knowledge their designers did not anticipate. Six of the agents discovered a technique—letting go of a block mid‑jump—to throw it farther, a behavior not explicitly programmed. Cognitive scientist Eunice Yiu of UC Berkeley noted the results are intriguing but said more evidence is needed to confirm genuine play. The researchers suggest that, like childhood, a play‑based regime might complement task‑driven training for future robotic control systems.