DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?
Vision-Language Models used for robot planning often waste compute by scaling test-time resources uniformly. DIRECT introduces a dynamic routing framework that uses scene context to decide which specific planner to use for each subtask. This approach shows that allocating compute based on task needs—rather than blindly increasing it—improves performance while significantly reducing latency and cost.