MA-VLCM: A Vision Language Critic Model for Value Estimation of Policies in Multi-Agent Team Settings

Researchers propose MA-VLCM, a framework that uses pretrained vision-language models as centralized critics in multi-agent reinforcement learning instead of learning critics from scratch. This approach significantly improves sample efficiency and enables zero-shot generalization while producing compact policies suitable for resource-constrained robots.

• →MA-VLCM replaces learned centralized critics in multi-agent reinforcement learning with fine-tuned pretrained vision-language models.
• →The framework eliminates the need for critic learning during policy optimization, significantly improving sample efficiency.
• →The approach produces compact execution policies suitable for deployment on resource-constrained robots.
• →Results demonstrate good zero-shot return estimation across different VLM backbones in both in-distribution and out-of-distribution scenarios.
• →The system addresses computational constraints in heterogeneous multi-robot systems with diverse embodiments.