Visual Dexterity

This is the codebase for Visual Dexterity: In-Hand Reorientation of Novel and Complex Object Shapes, accepted by Science Robotics. While we provide the code that uses the D'Claw robot hand, it can be easily adapted to other robot hands.

[Project Page]

:books: Citation

@article{chen2023visual,
    author = {Tao Chen  and Megha Tippur  and Siyang Wu  and Vikash Kumar  and Edward Adelson  and Pulkit Agrawal },
    title = {Visual dexterity: In-hand reorientation of novel and complex object shapes},
    journal = {Science Robotics},
    volume = {8},
    number = {84},
    pages = {eadc9244},
    year = {2023},
    doi = {10.1126/scirobotics.adc9244},
    URL = {https://www.science.org/doi/abs/10.1126/scirobotics.adc9244},
    eprint = {https://www.science.org/doi/pdf/10.1126/scirobotics.adc9244},
}

@article{chen2021system,
    title={A System for General In-Hand Object Re-Orientation},
    author={Chen, Tao and Xu, Jie and Agrawal, Pulkit},
    journal={Conference on Robot Learning},
    year={2021}
}

:gear: Installation

Dependencies

PyTorch
PyTorch3D
Isaac Gym (results in the paper are trained with Preview 3.)
IsaacGymEnvs
Minkowski Engine
Wandb

Download packages

You can either use a virtual python environment or a docker for training. Below we show the process to set up the docker image. If you prefer using a virtual python environment, you can just install the dependencies in the virtual environment.

Here is how the directory looks like:

-- Root
---- dexenv
---- IsaacGymEnvs
---- isaacgym

# download packages
git clone git@github.com:Improbable-AI/dexenv.git
git clone https://github.com/NVIDIA-Omniverse/IsaacGymEnvs.git

# download IsaacGym from: 
# (https://developer.nvidia.com/isaac-gym)
# unzip it in the current directory

# remove the package dependencies in the setup.py in isaacgym/python and IsaacGymEnvs/

Download the assets

Download the robot and object assets from here, and unzip it to dexenv/dexenv/.

Download the pretrained models

Download the pretrained checkpoints from here, and unzip it to dexenv/dexenv/.

Prepare the docker image

You can download a pre-built docker image:

docker pull improbableailab/dexenv:latest

Or you can build the docker image locally:

cd dexenv/docker
python docker_build.py -f Dockerfile

Launch the docker image

To run the docker image, you would need to have the nvidia-docker installed. Follow the instructions here

# launch docker
./run_image.sh # you would need to have wandb installed in the python environment

In another terminal

./visualize_access.sh
# after this, you can close it, just need to run this once after every machine reboot

:scroll: Usage

:bulb: Training Teacher

# if you are running in the docker, you might need to run the following line
git config --global --add safe.directory /workspace/dexenv

# debug teacher (run debug first to make sure everything runs)
cd /workspace/dexenv/dexenv/train/teacher
python mlp.py -cn=debug_dclaw # show the GUI
python mlp.py task.headless=True -cn=debug_dclaw # in headless mode

# if you wanna just train the hand to reorient a cube, add `task.env.name=DClawBase`
python mlp.py task.env.name=DClawBase -cn=debug_dclaw 

# training teacher
cd /workspace/dexenv/dexenv/train/teacher
python mlp.py -cn=dclaw
python mlp.py task.task.randomize=False -cn=dclaw # turn off domain randomization
python mlp.py task.env.name=DClawBase task.task.randomize=False -cn=dclaw # reorient a cube without domain randomization

# if you wanna change the number of objects or the number of environments
python mlp.py alg.num_envs=4000 task.obj.num_objs=10 -cn=dclaw

# testing teacher
cd /workspace/dexenv/dexenv/train/teacher
python mlp.py alg.num_envs=20 resume_id=<wandb exp ID> -cn=test_dclaw
# e.g. python mlp.py alg.num_envs=20 resume_id=dexenv/1d1tvd0b -cn=test_dclaw

:high_brightness: Training Student with Synthetic Point Cloud (student stage 1)

# debug student
cd /workspace/dexenv/dexenv/train/student
python rnn.py -cn=debug_dclaw_fptd
# by default, the command above used the pretrained teacher model you downloaded above, 
#if you wanna use another teacher model, add `alg.expert_path=<path>`
python rnn.py alg.expert_path=<path to teacher model> -cn=debug_dclaw_fptd

# training student
cd /workspace/dexenv/dexenv/train/student
python rnn.py -cn=dclaw_fptd

# testing student
cd /workspace/dexenv/dexenv/train/student
python rnn.py resume_id=<wandb exp ID> -cn=test_dclaw_fptd

:tada: Training Student with rendered Point Cloud (student stage 2)

# debug student
cd /workspace/dexenv/dexenv/train/student
python rnn.py -cn=debug_dclaw_rptd

# training student
cd /workspace/dexenv/dexenv/train/student
python rnn.py -cn=dclaw_rptd

# testing student
cd /workspace/dexenv/dexenv/train/student
python rnn.py resume_id=<wandb exp ID> -cn=test_dclaw_rptd

:rocket: Pre-trained models

We provide the pre-trained models for both the teacher and the student (stage 2) in dexenv/expert/artifacts. The models were trained using Isaac Gym preview 3.

# to see the teacher pretrained model
cd /workspace/dexenv/dexenv/train/teacher
python demo.py

# to see the student pretrained model
cd /workspace/dexenv/dexenv/train/student
python rnn.py alg.num_envs=20 task.obj.num_objs=10  alg.pretrain_model=/workspace/dexenv/dexenv/pretrained/artifacts/student/train-model.pt test_pretrain=True test_num=3 -cn=debug_dclaw_rptd