This is a github repository of a CALAMARI: Contact-Aware and Language conditioned spatial Action MApping for contact-RIch manipulation (CoRL 2023).
We trained with the GPU A6000 and ran inference on the RTX 3080 and RTX 2070.
conda create -n calamari python=3.8
conda activate calamari
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
conda env create -f environment.yml
We utilize heatmap extraction from Semantic Abstraction (Huy et al., CoRL 2022)."
git submodule add -f [email protected]:yswi/semantic-abstraction.git calamari/semantic_abstraction
pip install -e .
-
Download the dataset.zip from the link: https://www.dropbox.com/scl/fo/6w3p35agbu89ojp1mux5t/h?rlkey=0dxqegorjzo45tlzzy06y0w2z&dl=0
-
Make 'dataset' folder and upzip the dataset.
── calamari
│ ├── calamari
│ ├── dataset
│ │ │── wipe
│ │ │── sweep
│ │ │── push
│ ├── script
...
python script/train.py --task <TASK NAME> --logdir <FOLDER NAME> --gpu_id <GPU IDX>
Note: We use A6000 (48G) for training. You can decrease the batch size in config_multi_conv.py to match your GPU capacity, but a performance drop should be expected.
Generate heatmaps of the custom data.
python script/dataprocessing/generate_heatmap.py --task <TASK>
This repository trains the policy based on the RLbench dataset. RLbench code for inference and data collection will be released soon.