TransferLearning_DiffusionPolicy

Repository for the "Deep Learning Lab" offered by the University Freiburg. This project focusses on Transfer Learning with a learned Diffusion Policy.

See docs/README.md for unabridged install and run instructions

Setup

With Miniconda installed on your machine, execute the setup script:

# Allow file execution
chmod +x conda_env_setup.sh

# Execute setup script
./conda_env_setup.sh

# Collect Faucet Assets
python -m mani_skill2.utils.download_asset partnet_mobility_faucet

# Install 3D Diffusion Policy (Point Cloud)
cd ..
git clone https://github.com/YanjieZe/3D-Diffusion-Policy.git && cd 3D-Diffusion-Policy
cd 3D-Diffusion-Policy
pip install -e .
cd ../../TransferLearning_DiffusionPolicy

# Install pytorch3d
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"

You should now be in the activated conda environment, dlproject.

Workflow:

1. Collecting Demos

Download Demonstrations

# Download all rigid-body demonstrations (TODO: Download only necessary demos)
gdown https://drive.google.com/drive/folders/1pd9Njg2sOR1VSSmp-c1mT7zCgJEnF8r7 --folder -O demos/

After downloading you have to unzip it into the demos folder

Generate Training Data via Prerecorded Trajectories

• Place the downloaded trajectories (.h5 and .json) into the following file structure:
  • category_a will include any original training faucet models.
  • category_b will include any transfer faucet models (exclude if not needed)
• Feel free to change the experiment name, but do not change the names for the subdirectories (category_a and category_b).

demos
 ├── your_experiment          # Parent folder, specify as dataset path (change name, pass to train_3dp.py)
 |    ├── category_a          # First category of faucet, do not change this name
 |    |   └── *.h5, *.json    # Place any faucets from the first category in here (ex. 5000.h5, 5000.json)
 |    └── category_b          # First category of faucet, do not change this name
 |        └── *.h5, *.json    # Place any faucets from the second category in here

• Once you have added all desired models, run the following command at the root directory of the project:

# Generate demonstrations and save as .zarr
bash scripts/generate_data.sh "demos/your_experiment" 10 2

• Change the integer args to change number of trajectories to be collected for category_a and category_b respectively.
• This will save data into .zarr format.
• Pass this zarr file location demos/your_experiment/full_dataset.zarr into the training script (or hydra file in maniskill2_faucet.yaml at dataset.zarr_path).

2. Train Diffusion Policy

# 3D Diffusion
bash scripts/train_3dp.sh dp3 maniskill2_faucet 0322 0 0 your_experiment

3. Evaluate Task

# 3D Diffusion
bash scripts/eval_3dp.sh dp3 maniskill2_faucet 0322 0 0

Methodology:

• Single Faucet
  1. Single Faucet Original (without finetuning)

     demos
     ├── single_faucet_original          
     |    └── category_a          # 10 demos
     |        └── 5001  
     

  2. Single Faucet Transfer (without finetuning)

     demos
     ├── single_faucet_transfer         
     |    └── category_a             # 10 demos
     |        └── 5006   
     

  3. Single Faucet Finetuned

     demos
     ├── single_faucet_finetuned          
     |    ├── category_a             # 10 demos
     |    |   └── 5001    
     |    └── category_b             # 2 Demos
     |        └── 5006    
     

• Single Category
  1. Single Category Original (without finetuning)

     demos
     ├── single_category_original         
     |    └── category_a                     # 10 demos
     |        └── 5001, 5037, 5064, 5025   
     

  2. Single Category Transfer (without finetuning)

     demos
     ├── single_category_transfer         
     |    ├── category_a                     # 10 demos
     |    |   └── 5001, 5037, 5064, 5025   
     |    └── category_b                     # 10 demos
     |        └── 5006    
     

  3. Single Category Finetuned

     demos
     ├── single_category_finetuned         
     |    ├── category_a                     # 10 demos
     |    |   └── 5001, 5037, 5064, 5025   
     |    └── category_b                     # 2 demos
     |        └── 5006    
     

• Multi-Category
  1. Multi-Category Original (without finetuning)

     demos
     ├── multi_category_original         
     |    └── category_a                         # 10 demos
     |        └── 5001, 5037, 5064, 5025, 5006    
     

  2. Multi-Category Transfer (without finetuning)

     demos
     ├── multi_category_transfer         
     |    ├── category_a                         # 5 demos
     |    |   └── 5001, 5037, 5064, 5025, 5006    
     |    └── category_b                         # 5 demos
     |        └── 5005, 5053, 5028, 5052    
     

  3. Multi-Category Finetuned

     demos
     ├── multi_category_transfer         
     |    ├── category_a                         # 5 demos
     |    |   └── 5001, 5037, 5064, 5025, 5006    
     |    └── category_b                         # 1 demos
     |        └── 5005, 5053, 5028, 5052    
     

Setup:

• Robot: Franka Robot
• Observation input:
  • Robot State: joint position and tool control point (qpos+tcp)
  • Segmented pointcloud showing: robot, faucet
    • Point-cloud size: 1024
    • Point-cloud sampling method: Furthest-Point Sampling
• Epochs/Batchsize: 3000/128
• Control method: pd_joint_pos
• Segmented pointcloud: robot, faucet