Open Models Almanac
Methodsdata curationdata mixture & curriculum

single mention · data curation

Large-scale multi-source robot-manipulation pre-training mixture

A pre-training data mixture for robot manipulation models that combines newly collected dexterous manipulation data with large amounts of previously collected robot manipulation datasets.1 It serves as the base pre-training data in a framework where the model is later evaluated zero-shot or fine-tuned on downstream tasks.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. Our pre-training mixture consists of 10,000 hours of dexterous manipulation data from 7 different robot configurations and 68 tasks, in addition to large amounts of previously collected robot manipulation data from OXE [10], DROID [23], and Bridge [52 — π0\pi_{0}: A Vision-Language-Action Flow Model for General Robot Control Thanks: Physical Intelligence, San Francisco, California, USA.…
  2. consists of pre-training on highly diverse data, followed by either out-of-box evaluation or fine-tuning to complex downstream tasks — π0\pi_{0}: A Vision-Language-Action Flow Model for General Robot Control Thanks: Physical Intelligence, San Francisco, California, USA.…

Also called pre-training mixture.

Filed under data curation › data mixture & curriculum