ExpertGen Framework Automates Expert Policy Learning for Robotics via Sim-to-Real Transfer

The ExpertGen framework streamlines the process of learning expert policies in simulation, facilitating effective sim-to-real transitions for robotics. It tackles the issue of gathering extensive, high-quality robotics data, as acquiring human demonstrations via teleoperation is too costly to scale in real-world settings. ExpertGen begins with a behavior prior derived from a diffusion policy that has been trained on imperfect demonstrations, which can be generated by large language models or sourced from humans. Subsequently, reinforcement learning refines this prior to enhance task success by adjusting the initial noise of the diffusion model while maintaining the original policy's integrity. By freezing the pretrained diffusion policy, ExpertGen ensures exploration remains within safe, human-like behavior boundaries. This method allows for the efficient development of robust and generalizable behavior cloning policies. The framework was detailed in the paper titled "ExpertGen: Scalable Sim-to-Real Expert Policy Learning from Imperfect Behavior Priors," with the arXiv identifier 2603.15956v2. The announcement type is replace-cross, signifying an updated version on the arXiv preprint server.