Guiding AI learning… with AI?
Excited to share our paper “Palette Inpainting Diffusion Curriculum Reinforcement Learning (PIDCRL)”, accepted at WCCI 2026!
Curriculum Reinforcement Learning (CRL) helps agents learn faster by structuring tasks from easy → hard. We introduce PIDCRL — a diffusion-based CRL framework that converts trajectory heatmaps into candidate goals via a pretrained image-to-image diffusion model. To select meaningful goals, we employ several strategies, including averaging, Q-value scoring and a trainable reward-based mechanism.
The result?
- No hand-crafted heuristics
- Goals that are both achievable and challenging
- Strong performance across multiple maze environments
PIDCRL matches or outperforms 10 state-of-the-art CRL methods.
Excited about the intersection of diffusion models + reinforcement learning — lots of potential ahead!
