Measurable Task Representation Learning for Automatic Curriculum Generation

A new approach to automatic curriculum generation in reinforcement learning is proposed, addressing limitations of interpolation-based methods in non-Euclidean task spaces. The method, detailed in arXiv:2605.23372v1, introduces measurable task representation learning to enable automatic curriculum generation for complex navigation tasks. Traditional interpolation-based curriculum reinforcement learning (CRL) assumes a Euclidean task space with meaningful distance metrics, which fails in non-Euclidean contexts. The proposed technique learns a measurable representation of tasks, allowing the agent to automatically generate intermediate tasks between initial and target distributions without relying on predefined metrics. This advances CRL by extending automatic curriculum generation to more challenging environments where task similarity cannot be easily measured.