DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?

The article introduces DIRECT, a routing framework designed to optimize the allocation of test-time compute for Vision-Language Models (VLMs) in embodied planning tasks. It demonstrates that strategic compute allocation can enhance performance while reducing latency and resource usage, achieving up to 65% lower average latency compared to stronger models. This approach is significant for AI practitioners as it allows for more efficient deployment of embodied agents in real-world applications, balancing compute costs with performance gains across various scaling dimensions.