/test_chop

Primary LanguagePython

Learning to Use Chopsticks in Diverse Gripping Styles

Zeshi Yang, KangKang Yin, and Libin Liu

ACM Trans. Graph., Vol. 41, No. 4, Article 95. (SIGGRAPH 2022)

Table of Contents

  1. Requirements and Dependencies
  2. Installation
  3. Running demo
  4. Training

Requirements and Dependencies

  • mujoco_py (version:1.50.1.1)
  • PyGLM (version: 0.4.8b1)
  • torch (version:1.9.0+cu11)
  • pytorch3d (latest)
  • Everything else in requirements.txt

Installation

If you need help with installing the dependecies please refer to our installation guide in installation.md.

Running demo

To run our demo, run

python test.py --xml ./released_models/demo.xml --traj ./released_models/demo.txt --model_path ./released_models/model.tar

to visualize our pre-trained control policy for some chopsticks grasping tasks.

Training

To train policies with generated trajectories, run

python train.py --xml 'path_to_MJCF' --traj 'path_of_trajectories' --threads 'num_threads' --logdir 'path_to_save_models'

For example:

python train.py --xml ./data/hand_xml_grasp/demo.xml --traj ./data/traj/ --threads 16 --logdir ./results

You can modify the task_placeholder.py to generate your own training data. Run

python task_placeholder.py --chop_xml ./data/xml_templates/standard/0/chop_kin.xml --hand_xml ./data/xml_templates/standard/0/sim.xml --task_name 'your_task_name' --pose ./data/xml_templates/standard/0/standard.txt

to generate trajectories and MJCF file used for training. The generated motion files will store in ./data/your_task_name/ and the MJCF file will be in ./data/hand_xml_grasp/.