search

google-deepmind / open_x_embodiment

Apache License 2.0
794 stars 54 forks source link

(Further) Explanation of action space required #32

Open ham952 opened 9 months ago

ham952 commented 9 months ago

Can you please let me know what are the units for 7 parameters, for instance for Bridge dataset. Are you using fixed axis or current axis representation to show variations in subsequent frames ? More precisely, how can we traceback the exact values for each of the 7 parameters in workspace from an episode ?

kpertsch commented 9 months ago

Hi! We don't do any modifications of the original action spaces used in each of the contributed datasets. Please check the dataset's website or reach out to the authors of each dataset you are interested in to figure out the specifications they use for their dataset. We don't have any more detailed information beyond what's publicly available in the "description" fields of the TF dataset, see e.g. the TFDS catalogue: https://www.tensorflow.org/datasets/catalog/berkeley_autolab_ur5

For Bridge specifically, I think the actions are represented in a robot-anchored coordinate system, but if you need more precise information, please reach out to Homer Walke et al for the details!