DexFlyWheel was proposed in September 2025 by Peking University, Harbin Institute of Technology, and PsiBot, and the relevant research results were published in the paper "DexFlyWheel: A Scalable and Self-improving Data Generation Framework for Dexterous Manipulation", was accepted as Spotlight by NeurIPS 2025.

DexFlyWheel is a scalable data generation framework employing a self-improving loop to continuously enrich data diversity. The framework has two key design features: IL + Residual RL for generating human-like and diverse data. Specifically, IL and Residual RL, combined with policy unrolling and data augmentation, form a self-improving loop. In each iteration, the policy generates trajectories, which are then enhanced in increasingly diverse scenarios and subsequently fed into the next iteration. This loop creates a flywheel effect, progressively expanding data diversity, enhancing policy generalization capabilities, and evolving into a robust, generalizable data generation agent.