Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement
The paper proposes VERITAS, a framework for improving robot policies through inference-time policy steering and self-improvement using a generator-verifier system.
Proposing a new framework for inference-time policy steering and self-improvement for robots using a generator-verifier system
Keywords
Before reading this…
Applications
- →Robotics
- →Autonomous systems
To understand this paper, make sure you know these concepts first:
- Understanding of robot policiesfind papers →
- Concept of generator-verifier systemsfind papers →
Abstract
More Like ThisRobots deployed in the real world should learn from their experience and improve over time. This requires a mechanism of practicing and learning from feedback. In this paper, we propose VERITAS, a generator-verifier framework for generalist robot policies for inference-time policy steering and self-improvement. We use a pre-trained generalist robot policy as a ``generator'' and pair it with a gradient-free ``visual verifier'' that evaluates actions at inference time. This framework enables inference-time steering that improves policy performance without additional training. We demonstrate that inference-time verification consistently outperforms vanilla generalists without training on additional demonstration data. Additionally, we demonstrate that the verified rollouts provide effective supervision for offline policy improvement: policies fine-tuned on verified self-generated trajectories achieve consistent performance gains. Notably, we find that post-training with verified rollouts achieves comparable efficiency to expert demonstrations, while requiring no human interventions. Our results highlight inference-time verification as a practical and scalable mechanism for improving robotic policies during deployment.