You've Got a Golden Ticket: Improving Generative Robot Policies With A Single Noise Vector

The article presents a novel approach to enhancing pretrained generative robot policies by utilizing a constant initial noise vector, termed a "golden ticket," instead of sampling from a Gaussian distribution. This method, applicable to various diffusion and flow matching policies, demonstrated significant performance improvements across 46 out of 51 tasks, with success rates increasing by up to 55% in simulated environments and 28% in real-world scenarios. The approach requires no additional training or infrastructure, making it easily deployable, and the authors have released a codebase containing pretrained policies and golden tickets for further experimentation.