Robots can score highly on simulated tasks even if they mostly repeat behavior seen in examples. That makes it hard to tell whether a system can handle new situations or whether the benchmark resembles its demonstrations.
The researchers used k-nearest neighbors (kNN), a method that finds the most similar stored examples. For each current robot observation, it retrieved nearby demonstration states and reused the action sequences that followed them, without training a control policy. They also compared retrieval with vision-language-action (VLA) robot models.
Retrieval succeeded on 91.7% of LIBERO tasks and had the smallest performance drop under tested perturbations. This suggests benchmark scores can reflect matching demonstrations, not only new control skills, and that retrieval can help check how much a model generalizes.