griffinlabs/EgoDex-LeRobot-v3.0

EgoDex dataset from https://github.com/apple/ml-egodex, converted to LeRobot datasets v3.0 format.

This repo is structured with train / test directories at the top level, under each of which are multiple subdirectories, each corresponding to one directory in the original EgoDex data format. Each of these represents one LeRobot dataset, so to load the full EgoDex training set you have to load each subdirectory as its own LeRobotDataset and then combine them (e.g. w/ PyTorch's ConcatDataset), or do it automatically using MultiLeRobotDataset.

Confidence scores were not kept during the conversion.

Due to having a large number of features, it may take up large amounts of RAM to load the parquet tables. To avoid this, one could drop the unused features from the datasets using the lerobot-edit-dataset command.

Refer to EgoDex's GitHub repository for more information on the available features.

Split folders in training set

In the training set only, a few folders in the original EgoDex data format had to be split into smaller subsets in order to speed up the data conversion process:

• basic_fold
• basic_pick_place
• play_piano
• thread_unthread_bead_necklace
• vertical_pick_place

These folders are suffixed with -part* in our converted format, e.g. basic_fold-part1.