To improve the potential of Vision-Language Models (VLMs) as zero-shot goal-oriented value functions, we propose a zero-shot value function learning method called VITA via test-time adaptation. VITA updates lightweight adaptive modules via gradient descent to enhance generalization and temporal inference capabilities. During training, it employs a dissimilarity-based sampling strategy that selects semantically diverse trajectory segments. In real-world robotic manipulation tasks, VITA generalizes across diverse non-distributed tasks, environments, and implementations within a single training environment, outperforming state-of-the-art zero-shot methods using autoregressive VLMs. Furthermore, VITA's zero-shot value estimates can be used for reward shaping in offline reinforcement learning, generating multi-task policies that outperform policies trained with fuzzy logic dense rewards in simulation on the Meta-World benchmark.
