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

A research team from Brown University, MIT, and Google has introduced a remarkably simple yet effective technique for improving pre-trained generative robot policies. Their method, dubbed the 'Golden Ticket,' involves replacing the random noise vector typically sampled from a Gaussian distribution at the start of a policy's action sequence with a single, carefully optimized constant vector. This 'ticket' is found using a Monte-Carlo search that evaluates rollouts based on task rewards, all while keeping the original policy completely frozen. No new neural networks are trained, and the approach makes minimal assumptions, requiring only the ability to inject initial noise and calculate sparse rewards.

The impact is substantial. The team demonstrated improvements in 38 out of 43 tested robot manipulation tasks across simulation and real-world benchmarks. Success rates increased by up to 58% for some simulated tasks and by 60% within just 50 search episodes for real-world deployments. The method is broadly applicable to diffusion policies, flow matching policies, and the Vision-Language-Action models that often use them. Beyond single-task improvement, the researchers found that different 'Golden Tickets' can produce a diverse set of behaviors, naturally defining a Pareto frontier for balancing objectives like speed and accuracy. They also observed positive transfer, where a ticket optimized for one task could boost performance in related tasks.

The researchers have released a codebase containing pre-trained policies and the discovered 'Golden Tickets' for several simulation benchmarks. This work provides a powerful, low-cost tool for fine-tuning the behavior of complex robotic AI systems post-training, offering significant performance gains without the computational expense of full model retraining or fine-tuning.