search

real-stanford / diffusion_policy

[RSS 2023] Diffusion Policy Visuomotor Policy Learning via Action Diffusion
https://diffusion-policy.cs.columbia.edu/
MIT License
1.09k stars 205 forks source link

What are the differences between low_dim.hdf5 and low_dim_abs.hdf5? #35

Open Junxix opened 7 months ago

Junxix commented 7 months ago

Hi Cheng, I'm wondering what are the differences between low_dim.hdf5 and low_dim_abs.hdf5 since the dataset in the robomimic environment only has low_dim.hdf5 and image.hdf5. How is the low_dim_abs.hdf5 and image_abs.hdf5 dataset produced? Thanks for your help!

cheng-chi commented 6 months ago

Hi @Junxix, in robomimic's offical dataset, all actions are represented as the delta with respect to the current pose. We call this action represetnation "velocity control" in this paper. Unfortunately, CNN-based diffusion policy performs very poorly under this action represetantion. Therefore, we converted all robomimic datasets into absolute posititional actions using robomimic_dataset_conversion.py. The result of this conversion is saved to low_dim_abs.hdf5 and image_abs.hdf5 respectively.

deavn2236 commented 1 month ago

The reading method for the diffusion policy dataset is in the form of zarr. Why is it HDF5? Do we need to convert HDF5 to zarr?