Systematic selection of quantile sampling strategies

Develop a systematic strategy-selection method for choosing among quantile-based action sampling strategies for vision-language-action robot control, given that different strategies produce task-dependent gains and losses.

Background

The Quantile Head predicts ordered marginal action quantiles in a single forward pass. These quantiles can be decoded deterministically through the median or stochastically through alternative sampling procedures, including different sampling windows, temporal correlation structures, and density-weighted proposals.

Experiments show that sampling can improve performance in some settings but degrade it in others. For example, a narrow sampling window improves LIBERO-Long performance, whereas the controlled obstacle experiment yields low and highly variable success across sampling strategies. The paper therefore identifies the need for a systematic method to select an appropriate sampling strategy rather than relying on manually chosen heuristics.