Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement

The paper introduces VERITAS, a generator-verifier framework designed for generalist robot policies that facilitates inference-time steering and autonomous policy improvement. By utilizing a pre-trained robot policy as a generator and a gradient-free visual verifier for action evaluation, the framework enhances policy performance without additional training, demonstrating superior results compared to traditional methods. This approach allows for effective offline policy improvement using verified self-generated trajectories, offering a scalable solution for enhancing robotic capabilities during real-world deployment.