1. Build the Environment

• Run the following command in your terminal where the environment.yml file is located:

  conda env create -f environment.yml

• Then, activate your new environment:

  conda activate Touch3D

2. Install PyTorch and Related Packages

• Now that your environment is set up with Python 3.9, run the following command to install PyTorch, torchvision, torchaudio, and the appropriate CUDA toolkit. Below is for CUDA 12.1:

  conda install pytorch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 pytorch-cuda=12.8 -c pytorch -c nvidia

3. (Optional) Install Additional pip Packages

• If you didn’t include the pip requirements in the YAML file, you can install them now:

  pip install -r requirements.txt

4. (Optional) Training the Network

• Create the folder 'Training/Logs' (the checkpoints and tf events will be saved here)

• Run the following command in your terminal to train the network:

  python PPO.py

  Note: The network will train from scratch for 500k steps. You can change the parameters in PPO.py file or give a pretrained model path in conf/RL.yaml

5. Test the Network on Unseen Objects

• Create the folder 'Outputs' (3D poit cloud will be saved here)

• Set configuration at conf/test.yaml for testing:

  • RL:

    • pretrain_model_path: Training/Logs/PPO_Contact_AMB/your_latest_model.zip

  • Environment:

    • Object:
      • object_name: Name of the test object
      • urdf path of the object: objects/ycb/object/model.urdf

• Run the following command in your terminal to test the model:

  python test.py

The testing will begin and point cloud is generated after 5000 evaluation steps are over or the sensor goes out of bounds.

**Note:** If due to some reason the point cloud cannot be visualized after testing is finished; follow the steps below to visualize the generated point cloud.

6: (optional) Point Cloud Visualization

• Navigate to the visualize_npy.py file

• Change the default path of the point cloud to the .npy file generated during testing, Example: outputs/DEMO_model_96.95.npy

• Execute the file:

  python visualize_npy.py

  NOTE:The live visualiztion is turned off by default to improve performance

7: Results

• This is the live visualization of testing the CNN policy with PPO on unseen 3D object(strawberry and mustard can in this particular case). The test is conducted for a single episode of 5000 steps.

• The point cloud visualization of complex objects is shown below: