Robot manipulation policies face a trade-off between speed and accuracy. Computationally lightweight policies can predict executable actions with low inference latency, but their prediction errors can limit task success. On the other hand, iterative generative policies tend to achieve higher task success rates, though the abstract notes they are more computationally demanding — the text cuts off before detailing the exact cost.

The paper, titled "Refine Actions If Needed," appears to address this trade-off by proposing an adaptive approach: refine actions only when necessary, rather than always or never. This would let a fast base policy act immediately when its predictions are good, while triggering a more expensive refinement process when errors are likely.

Because the abstract is truncated, the source does not include experimental results or implementation details. The contribution, as stated, is the conceptual strategy of conditionally applying refinement to balance latency and task success in robot manipulation.