Skip to content

Latest commit

 

History

History
110 lines (64 loc) · 3.4 KB

Download.md

File metadata and controls

110 lines (64 loc) · 3.4 KB
Raw

Dataset Download

All datasets and pretrained weights are stored in a box folder at https://rutgers.box.com/s/uzozemx67kje58ycy3lyzf1zgddz8tyq. These files are located in datasets and weights folders.

It takes a lot of time to download each file from browser manually. Therefore, I prepare a single compressed file all_datasets_and_weights.tar.gz that includes all files (in that box folder). You can choose to just download that and extract it. It will extract to a folder called release and you can find all files inside.

Push-T

Directly download from the origin source.

cd pusht
mkdir data && cd data
wget https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip
unzip pusht.zip && rm -f pusht.zip && cd ..

ALOHA

cd aloha
mkdir data && cd data

# Download datasets/aloha_human_demo_with_waypoints.zip
# from the box folder, and put it in this data folder

unzip aloha_human_demo_with_waypoints.zip 
mv aloha_human_demo_with_waypoints/lerobot .
rm -f aloha_human_demo_with_waypoints.zip aloha_human_demo_with_waypoints && cd ..

This dataset shares the same episodes with the original one but adds 2d waypoints. You can use the script aloha/compute_waypoints.py to re-generate this dataset from origin lerobot data.

RLBench

cd rlb
mkdir data && cd data

# Download datasets/RLBench.tar
# from the box folder, and put it in this data folder

tar xvf RLBench.tar
rm -f RLBench.tar && cd ..

This dataset contains the original pre-generated RLBench train / test demonstration from peract. However, it is much smaller (only 6GB vs hundreds of GBs). Therefore, it is much easier to get started with.

The reason is that I only keep the key frames from the original dataset. The back-story is:

  1. RLBench has a key frame extraction procedure, see keypoint_discovery function in rlb/dataset.py. Many existing works use this code snippet.
  2. There has been a long-standing "bug", regarding data sampling, in existing works. This "bug" significantly increases the sampling ratio on key frames. Read more into this issue here: peract/peract#6 (comment).
  3. Based on my personal experience, I found only key-frames contribute to the learning of the policy. Therefore, I simplify the implementation ("fix" this "bug") and trim the training set.
  4. I optimize the code a little bit so evaluation do not read the full testing episodes, in doing so, the test set is also trimmed.

Pretrained Weights

Here only the instructions on downloading the models of our main results are provided. The training logs of other experiments are stored in the training-logs folder (in box). Those experiments are detailed in the Experiments.md

Push-T

cd pusht
mkdir weights

# Download weights/pusht/epoch=2000-test_mean_score=0.865.ckpt from Box
# and put it in this weights folder

cd ..

There are better ones, but I forget to save them...

ALOHA

cd aloha
mkdir weights

# download models/aloha/model.transfer_cube.safetensor and models/aloha/model.insertion.safetensors
# from the box folder, and put them in this weights folder

cd ..

RLBench

cd rlb
mkdir weights

# download  models/rlb/arp_model_80000.pth and models/rlb/arp_plus_model_70000.pth
# from the box folder, and put them in this weights folder

cd ..