Open Models Almanac
Methodsdata curation

single mention · data curation

Past/future action-window sampling

Past/future action-window sampling is a preprocessing step that constructs each training sample from a window of low-dimensional action timesteps centered on an anchor timestep within an episode.1 Past timesteps let the model condition on recent action history and proprioceptive context, while multiple future timesteps enable temporal action chunking.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. each training sample is constructed around an anchor timestep within an episode. The low-dimensional window centered at — VLA Foundry: A Unified Framework for Training Vision-Language-Action Models
  2. Including past timesteps allows the model to condition on recent action history and proprioceptive context; predicting multiple future timesteps allows for temporal action chunking — VLA Foundry: A Unified Framework for Training Vision-Language-Action Models

Also called Past/Future Action Window.

Filed under data curation