Haonan Chen¹, Cheng Zhu¹, Yunzhu Li², Katherine Driggs-Campbell¹

¹University of Illinois, Urbana-Champaign, ²Columbia University,

• UR5-CB3 or UR5e (with RTDE Interface)
• Alternative: Kinova Gen 3

• Sensors:
  • 2× Intel RealSense D415
  • USB-C cables and mounting hardware
• Control Interfaces:
  • 3Dconnexion SpaceMouse (teleoperation)
  • GELLO Controller (teleoperation)
• Custom Components:

  • 3D Printed Hammer and Nail

    To install a tool on the UR5E, we provide two types of fast tool changers:

    • 3D-Printed Clipper

      • Requires a connector to attach tools to the Clipper.
        • Example: The hammer linked above already includes the connector.

      The upper Clipper is connected to the Clipper base using one M4×16 screw and one M4 nut. A clipper gasket is provided to place between the UR5E robot and the Clipper. If the gasket is chosen to use, you should use four M6×30 screws. Without the clipper gasket, four M6×24 screws will work as well. Four M6 screw gasket will be used in both conditions.

    • 3D Printed Mounter

      • Suitable for both 3D-printed and standard tools.
      • Secured using one or two 3D-printed screws.

      To connect the Mounter with UR5E robot, you should use four M6×12 screws. Four M6 screw gasket will be used here to make it tightly connected.

We recommend using Mambaforge over the standard Anaconda distribution for a faster installation process. Create your environment using:

1. Install the necessary dependencies:

   sudo apt install -y libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf libglm-dev

2. Clone the repository:

   git clone --recursive https://github.com/Tool-as-Interface/Tool_as_Interface.git
   cd Tool_as_Interface/
   git clone https://github.com/xinyu1205/recognize-anything.git third_party/Grounded-Segment-Anything/recognize-anything

3. Update mamba, and create and activate the environment:

   To update mamba and create the environment, use the following commands:

   mamba install mamba=1.5.1 -n base -c conda-forge
   mamba env create -f conda_environment_real.yml
   mamba activate ti

4. Install packages:

   # grounded sam
   export AM_I_DOCKER=False
   export BUILD_WITH_CUDA=True
   export CUDA_HOME=/usr/local/cuda-11.8
   pip install https://artifactory.kinovaapps.com:443/artifactory/generic-public/kortex/API/2.6.0/kortex_api-2.6.0.post3-py3-none-any.whl
   pip install --no-build-isolation -e third_party/Grounded-Segment-Anything/GroundingDINO
   pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.0.0/index.html
   pip install third_party/Grounded-Segment-Anything/grounded-sam-osx/transformer_utils
   # FoundationPose
   CONDA_ENV_PATH=$(conda info --base)/envs/$(basename "$CONDA_PREFIX")
   EIGEN_PATH="$CONDA_ENV_PATH/include/eigen3"
   export CMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH:$EIGEN_PATH"
   cd third_party/FoundationPose
   CMAKE_PREFIX_PATH=$CONDA_PREFIX/lib/python3.10/site-packages/pybind11/share/cmake/pybind11 console build_all_conda.sh
   cd ../..

5. Download the checkpoints

   bash setup_downloads.sh

To ensure everything is installed correctly, run the following commands:

python -c 'import torch; print(torch.__version__); print(torch.cuda.is_available())'
python -c 'import torchvision; print(torchvision.__version__)'
python -c "from groundingdino.util.inference import Model; from segment_anything import sam_model_registry, SamPredictor"

mkdir -p data

Place the hammer_human dataset in the data folder. The directory structure should be:

To obtain the dataset, download the corresponding zip file from the following link:
hammer_human dataset

Once downloaded, extract the contents into the data folder.

Directory structure:

data/
└── hammer_human/

Activate conda environment and login to wandb (if you haven't already).

conda activate ti
wandb login

Adjust the arguments in the calibrate_extrinsics function located in ti/real_world/multi_realsense.py to calibrate multiple cameras. The calibration is performed using a 400×300 mm Charuco board with a checker size of 40 mm (DICT_4X4).

The robot_base_in_world parameter is manually measured and tuned in ti/real_world/multi_realsense.py.

Run the following command to perform the calibration:

python ti/real_world/multi_realsense.py

Start the demonstration collection script. The following script applies to both human play collection and robot demonstration collection.

• Press C to start recording.
• Press S to stop recording.
• Press Backspace to delete the most recent recording (confirm with y/n).

Run the following script to collect human video demonstrations:

python scripts/collect_human_video.py

Once the human play data has been collected, process the raw data using:

python scripts/preprocess_human_play.py

If you want to use GELLO, please calibrate the GELLO offset using the following script:

python ti/devices/gello_software/scripts/gello_get_offset.py

After calibration, update the YAML configuration file:

ti/devices/gello_software/gello.yaml

For details on performing the calibration, refer to:

ti/devices/gello_software/README.md

Once the calibration is complete, update the argument in scripts/demo_real_ur5e.py to select either Spacemouse or GELLO, then run the following command to collect robot demonstration data:

python scripts/demo_real_ur5e.py

To launch training, run:

python scripts/train_diffusion_policy.py \
  --config-name=train_diffusion_policy.yaml 

Modify eval_real_ur5e_human.py or eval_real_ur5e.py, then launch the evaluation script:

• Press C to start evaluation (handing control over to the policy).
• Press S to stop the current episode.

For eval_real_ur5e_human.py, we assume that the tool and the end-effector (EEF) are rigidly attached. Therefore, the tool pose estimation only needs to be performed once.

• Press T once to estimate the tool's pose in the EEF frame.

python eval_real_ur5e_human.py 

python eval_real_ur5e.py 

If you encounter the following error:

AttributeError: module 'collections' has no attribute 'MutableMapping'

Resolve it by installing the correct version of protobuf:

pip install protobuf==3.20.1

• Policy training implementation is adapted from Diffusion Policy.