Scaling test-time compute is not free in robotics: in a hierarchical stack the planner is invoked repeatedly, so per-call costs from chain-of-thought reasoning, larger models, or denser memory history compound and can multiply base model latencies. Cutting compute uniformly mitigates latency but risks failure on harder tasks. We therefore study where compute actually yields gains across three axes—reasoning depth, model size, and memory history—motivating our framework.
Understanding the Compute Allocation Problem
In embodied planning, a robot must decide when and where to allocate additional computation during deployment. Unlike in static domains like language modeling or game playing, robotics imposes real-time constraints: each planning call consumes time and energy, and the planner is called many times throughout a task. Uniformly increasing compute improves performance on difficult scenarios but wastes resources on easy ones, while uniformly reducing compute saves resources but may cause failures on challenging edge cases. The central question is whether selective allocation—applying more compute only where it provides measurable benefit—can improve overall efficiency and reliability.
Three Axes of Test-Time Compute
The paper investigates three distinct dimensions along which test-time compute can be scaled in embodied planners. First, reasoning depth refers to how many steps or tokens the model generates before committing to an action—longer chain-of-thought reasoning may yield better plans but adds latency. Second, model size captures the capacity of the underlying neural network; larger models typically perform better but are slower and more expensive. Third, memory history concerns how much of the agent's past experience is retained and used for decision-making, with denser histories providing more context at the cost of increased processing time. Understanding the interaction between these axes is critical for designing efficient embodied planning systems.
The DIRECT Framework
To systematically study where test-time compute yields the greatest returns, the authors introduce DIRECT, a framework that evaluates the marginal benefit of increasing compute along each axis for individual planning queries. The approach works by measuring how much performance improves when a planner is given more compute, and then learning when to invoke higher-cost variants. A key insight is that not all queries benefit equally from additional compute: some are already solvable with minimal resources, while others require the full capacity of larger models or deeper reasoning. By routing each query to the most appropriate compute level, the framework aims to match resource allocation to task difficulty.
Empirical Findings Across Axes
The experimental results reveal several patterns. For reasoning depth, longer chain-of-thought reasoning provides diminishing returns on simple navigation and manipulation tasks but remains valuable for complex multi-step tasks. Larger models consistently outperform smaller ones, but the performance gap is concentrated on tasks requiring spatial reasoning or precise geometric understanding. Memory history shows a mixed profile: denser histories help in partially observable environments but add overhead without benefit in fully observable settings. Critically, the benefits of scaling along different axes are not independent—combining deeper reasoning with larger models can yield super-linear gains on the hardest tasks.
Practical Implications for Robotics
The findings have direct implications for robotic system design. Rather than deploying a single fixed-size planner, engineers can build adaptive systems that allocate compute dynamically based on task difficulty. For example, easy pick-and-place operations might use a small model with shallow reasoning, while complex assembly tasks would trigger larger models with deeper chain-of-thought processing. The DIRECT framework provides a principled way to identify these thresholds and build the routing logic. On physical robot hardware, this adaptive approach demonstrated improved task completion rates while maintaining acceptable latency compared to uniformly high-compute baselines.
Frequently Asked Questions
How does test-time compute differ from training-time compute in robotics? Test-time compute refers to the resources used during deployment when the robot must make decisions in real time, while training-time compute is used during model development and optimization. In robotics, test-time compute is more constrained by latency and energy budgets.
Which axis of compute scaling provides the most consistent gains across tasks? Model size shows the most consistent performance improvements across diverse tasks, though with diminishing returns. The gains from deeper reasoning and denser memory are more task-dependent.
Can the DIRECT framework be applied to any embodied planner? The framework is designed to be planner-agnostic and can be applied to any hierarchical embodied system where the planner is called repeatedly, though the specific compute thresholds may need to be tuned per domain.
Does adaptive compute allocation increase system complexity significantly? While it adds a routing component, the overhead is minimal compared to the potential savings from avoiding unnecessary compute on easy tasks and preventing failures on hard tasks.