DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?
Researchers have developed a new framework called DIRECT to optimize the allocation of test-time compute for embodied AI planners. This method uses multimodal scene context to dynamically route compute, improving performance-cost trade-offs compared to fixed model selection. Experiments on simulated and physical robotic tasks demonstrate that DIRECT can achieve comparable success rates to stronger models at significantly lower latency and cost. AI
IMPACT Optimizes resource allocation for embodied AI, potentially enabling more efficient deployment of robotic systems.