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

Researchers introduce DIRECT, a routing framework that intelligently allocates computational resources at test-time for Vision-Language Models used in embodied AI planning. The system selectively chooses when to deploy expensive scaling strategies (deeper reasoning chains, larger models, expanded memory), achieving up to 65% lower latency than baseline approaches while maintaining or exceeding performance on robotic manipulation tasks.

The deployment of frontier AI models in robotics faces a fundamental constraint: scaling test-time compute uniformly across all decisions wastes resources and introduces latency that degrades real-world utility. DIRECT addresses this by routing individual planning prompts to different computational configurations based on scene context, recognizing that not all embodied decisions require equivalent computational investment. This approach reflects a maturing understanding in AI systems that raw compute scaling provides diminishing returns without intelligent allocation mechanisms.

The research builds on the trend of using Vision-Language Models as high-level planners for robotic agents, a shift that improves generalization but introduces new deployment challenges. Previous work scaled test-time compute indiscriminately, treating all planning decisions identically despite their varying complexity requirements. DIRECT's multimodal routing mechanism enables dynamic resource allocation, creating a more efficient capability-cost frontier.

For the robotics and embodied AI industry, this work directly impacts deployment feasibility. Reducing latency by 65% while maintaining performance changes the economics of real-world robotic applications, enabling deployment in latency-sensitive environments where frontier models previously proved impractical. The finding that different scaling axes (reasoning depth, model size, memory) produce qualitatively distinct capability gains suggests future optimization requires axis-specific routing rather than one-size-fits-all scaling strategies.

Looking forward, the validation on physical robotic systems demonstrates immediate practical relevance. Future developments likely involve learned routing policies that improve over time and cross-task optimization that leverages patterns from diverse embodied tasks. This positions intelligent compute allocation as a key enabler for practical AI robotics deployment.

• →DIRECT framework reduces latency by up to 65% while matching or exceeding performance of larger models through context-aware compute allocation
• →Different test-time scaling axes (chain-of-thought, model size, memory) produce qualitatively distinct capability gains requiring selective deployment
• →Uniform test-time compute scaling in embodied planning wastes resources without proportional performance improvements across all decision types
• →Physical validation on Franka arm demonstrates practical feasibility for real-world robotic manipulation and long-horizon task chaining
• →Intelligent routing mechanisms represent a necessary evolution beyond naive scaling for deploying frontier AI models in latency-sensitive robotic applications