Curiosity-Diffuser: Enhancing Robotic Policy Reliability via Curiosity-Guided Diffusion

A new research paper on arXiv proposes Curiosity-Diffuser, a method to improve the reliability of robotic policies by guiding conditional diffusion models to generate trajectories with lower curiosity. The approach uses a Random Network Distillation (RND) curiosity module to assess alignment with training data, then minimizes curiosity via classifier guidance to reduce overgeneralization. The paper also introduces a computationally efficient metric for evaluating policy reliability based on similarity between generated and training trajectories. The work addresses the instability of neural network models in robotic intelligence, which can lead to hallucinations and unsafe behaviors in real-world applications. The paper is available at arXiv:2503.14833.