Title:	Environment-Agnostic Autotelic Goal-Conditioned Reinforcement Learning
DNr:	Berzelius-2026-222
Project Type:	LiU Berzelius
Principal Investigator:	Jacek Malec <jacek.malec@cs.lth.se>
Affiliation:	Lunds universitet
Duration:	2026-08-18 – 2026-12-01
Classification:	10210
Homepage:	https://portal.research.lu.se/sv/projects/reinforcement-learning-and-curiosity-driven-exploration-for-robot/
Keywords:

Abstract

This continuation project extends ongoing research on environment-agnostic autotelic goal-conditioned reinforcement learning (GCRL). The project develops reinforcement learning agents capable of autonomously generating goals and acquiring transferable skills across multiple environments. Building on progress achieved during the initial allocation period, the continuation focuses on rigorous large-scale evaluation, hyperparameter optimization, ablation studies, and reproducibility validation of newly developed methods. Initial Berzelius-supported experiments have produced promising and potentially publishable results, including results that equal or exceed supervised state-of-the-art solutions under stricter constraints on multiple environments. The project has developed a family of autonomous goal-selection methods based on environment-agnostic goal representations, including flow-matching and discriminator-guided approaches. Across the current set of eight evaluation environments, several unsupervised variants achieve performance comparable to supervised baselines while avoiding reliance on task-specific supervision and associated out-of-distribution target issues. The project remains closely aligned with WASP research objectives in autonomous learning and artificial intelligence. Planned publication targets include leading reinforcement learning, robotics, and machine learning venues during autumn 2026.