luciferdarkstar / hcg

Cluttered-Scene 6D Grasping with Latent Plans

Hierarchical Policies for Cluttered-Scene Grasping with Latent Plans

[Website, Paper]

Installation

git clone https://github.com/liruiw/HCG.git --recursive

0. Setup: Ubuntu 16.04 or above, CUDA 10.0 or above, python 2.7 / 3.6

1. • (Training) - Install OMG submodule and reuse conda environment.
   • (Demo) - Install HCG inside a new conda environment

     conda create --name hcg python=3.6.9
     conda activate hcg
     pip install -r requirements.txt
     

2. Install PointNet++

3. Download environment data with bash experiments/scripts/download_data.sh

Pretrained Model Demo and Testing

0. Download pretrained models with bash experiments/scripts/download_model.sh
1. Demo model test bash experiments/scripts/test_demo.sh

Save Data and Offline Training

0. Download example offline data with bash experiments/scripts/download_offline_data.sh The .npz dataset (saved replay buffer) can be found in data/offline_data and can be loaded for training.
1. To train offline latent plan embedding and low-level policy bash ./experiments/scripts/train_embedding_ray_offline.sh bc_embedding_low_level.yaml BC or without ray. Add pretrained model after the policy for continual training and finetuning.
2. To train offline joint embedding and latent plan sampler bash ./experiments/scripts/train_embedding_ray_offline.sh bc_sampler_restep.yaml BC
3. To generate and save offline dataset bash ./experiments/scripts/train_ray_online_save_data.sh dqn_save_buffer.yaml DQN_HRL.

Online HRL Training

0. We use ray for parallel rollout and training. The training scripts might require adjustments based on the local machine. See config.py for some notes.
1. To train the latent critic online bash ./experiments/scripts/train_sampler_ray_online.sh dqn_critic_hrl.yaml DQN_HRL MODEL. Use visdom and tensorboard to monitor.

Testing

0. To test embedding and low-level policy bash experiments/scripts/test_ycb.sh demo_model_clutter .

Example 1	Example 2

1. To test high-level sampler and critic bash experiments/scripts/test_ycb_gen_12_mpc.sh demo_model_clutter

Example 1	Example 2

2. Testing only works with single GPU. Replace demo_model with trained model name. Logs and videos would be saved to output_misc

Note

0. Checkout core/test_realworld_ros_clutter.py for an example of real-world usages.
1. Related Works (OMG, GA-DDPG, ACRONYM, Unseen-Clustering).
2. To use the full Acronym dataset with Shapenet meshes, please follow ACRONYM to download the meshes and grasps and follow OMG-Planner to process and save in /data. filter_shapenet.json can then be used for training.
3. Please use Github issue tracker to report bugs. For other questions please contact Lirui Wang.

File Structure

├── ...
├── HCG
|   |── data  # training data
|   |   |── grasps      # grasps of the objects
|   |   |── objects     # object meshes, sdf, urdf, etc
|   |   |── robots      # robot meshes, urdf, etc
|   |   |── hcg_scenes  # test scenes
|   |   └── ...
|   |── env # environment
|   |   |── env_planner      # interface for the expert planner
|   |   |── panda_cluttered_scene     # environment and task
|   |   └── panda_gripper_hand_camera      # franka panda with gripper and camera
|   |── OMG # expert planner submodule
|   |── experiments
|   |   |── config      # hyperparameters for training, testing and environment
|   |   |── scripts     # runable scripts for train/test/validate
|   |   |── model_spec     # network architecture spec
|   |   |── cfgs # experiment config
|   |   └── object_index      # object indexes used for experiments
|   |── core    # agents and learning
|   |   |──  train_online      # online training
|   |   |──  train_offline     # offline training
|   |   |──  test_offline     # testing and video recording
|   |   |──  network      # network architecture
|   |   |──  agent        # learner and loss
|   |   |──  replay memory       # replay buffer
|   |   |──  trainer       # ray-related training
|   |   |──  test_realworld_ros_clutter    # example real-world scripts
|   |   └── ...
|   |── output  # trained model
|   |── output_misc  # log and videos
|   └── ...
└── ...

Citation

If you find HCG useful in your research, please consider citing:

@inproceedings{wang2021hierarchical,
author    = {Lirui Wang, Xiangyun Meng, Yu Xiang, and Dieter Fox},
title     = {Hierarchical Policies for Cluttered-Scene Grasping with Latent Plans},
booktitle = {IEEE Robotics and Automation Letters (RAL)},
year      = {2022}
}

License

The HCG is licensed under the MIT License.