Uber's 2019 POET research explores how a learning system can generate its own training challenges. In its simulated walking experiments, each obstacle course is paired with an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → learning to navigate it. Existing courses produce variations, and the system admits challenges that are neither too easy nor too difficult for its current learners.
POET also tests whether an agent trained on one course performs better on another, replacing the latter's learner when transfer helps. This combination of challenge generation and agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → transfer creates multiple developing curricula within one run. It offers a concrete approach to discovering useful training sequences without specifying every stepping stone in advance.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
