Before running the cup arrangement policy on other unseen cups, we suggest testing a standard cup (a 3d model that has similar shape to the ones in the collected data). Although this 3d printed cup is also unseen in the training data, it has the closest shape compared to a random cup.
PLA, 100% infill, tree support
Color: Sky blue color is recommended; other colors should work as well (white, black, grey are tested)
To purchase ARX robot arms, please visit their official website https://arx-x.com/
https://github.com/real-stanford/detached-umi-policy
git clone [email protected]:real-stanford/detached-umi-policy.git
cd detached-umi-policy
mkdir data && cd data && mkdir models && mkdir experiments
Download checkpoint and put it into data/models
Install python environment: recommend using mamba to create environments. Usage is the same as conda.
mamba env create -f conda_environment.yaml
conda activate umi
Test whether diffusion policy is successfully installed
python detached_policy_inference.py -i data/models/cup_wild_vit_l_1img.ckpt
After PolicyInferenceNode is listening on 0.0.0.0:8766, the policy inference process is successfully set up. Keep it running in the background when running umi code.
https://github.com/real-stanford/arx5-sdk/tree/yihuai
git clone [email protected]:real-stanford/arx5-sdk.git
cd arx5-sdk
Install conda environment for sdk compilation (python 3.10 is tested, while other versions may work as well)
mamba env create -f conda_environments/py310_environment.yaml
conda activate arx-py310