Researchers have developed Evolved Policy Gradients (EPG), a metalearning approach that evolves the loss function of learning agents. This method enables fast training on novel tasks and allows agents to adapt to new situations outside their original training regime. For example, an agent trained with EPG can learn to navigate to an object placed in a different location than it was during training. --- Why it matters: This matters because it could enable AI systems to generalize better across different scenarios and adapt more efficiently to new tasks, which is crucial for real-world applications where environments are often changing or unknown. Source: https://openai.com/index/evolved-policy-gradients