Crossway Diffusion: Improving Diffusion-based Visuomotor Policy via Self-supervised Learning
Paper • 2307.01849 • Published
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
This is the dataset repo for real world experiments presented in ICRA'24 paper Crossway Diffusion: Improving Diffusion-based Visuomotor Policy via Self-supervised Learning.
These datasets are collected in the same format as robomimic. To train on this dataset (and your own datasets), please modify your config file as follows:
dataset_path to the path of hdf5 file.shape_meta.shape_meta:
action:
shape:
- 4
obs:
pos:
shape:
- 4
type: low_dim
camera0:
shape:
- 3
- 120
- 160
type: rgb
camera1:
shape:
- 3
- 120
- 160
type: rgb
For the training and evaluation code, please visit our github repo.
Watch 🎬 video presentation!
For questions, please raise a github issue.