aawizard/graff

Repository for Learning Dexterous Grasping with Object-Centric Visual Affordances [ICRA 2021]

GRAFF

This repository contains the source code for the paper Learning Dexterous Grasping with Object-Centric Visual Affordances. Accepted at International Conference on Robotics and Automation (ICRA), 2021

Citing this work

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

@inproceedings{mandikal2021graff,
  title = {Learning Dexterous Grasping with Object-Centric Visual Affordances},
  author = {Mandikal, Priyanka and Grauman, Kristen},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2021}
}

Demo

To visualize the provided trained model, set up the environment correctly (see below) and run:

bash scripts/demo.sh

This will save videos of robot trajectories in expts/graff_trained/videos_stability. You can change the object as needed.

Overview

GRAFF is a deep RL dexterous robotic grasping policy that uses visual affordance priors learnt from humans for functional grasping. Our proposed model, called GRAFF for Grasp-Affordances, consists of two stages. First, we train a network to predict affordance regions from static images. Second, we train a dynamic grasping policy using the learned affordances. The key upshots of our approach are better grasping, faster learning, and generalization to successfully grasp objects unseen during policy training.

Installation

Mujoco

1. Download MuJoCo v2.0 binaries from the official website: https://www.roboti.us/download.html.
2. Download the activation key from here: https://www.roboti.us/license.html
3. Unzip the downloaded mujoco200 directory into ~/.mujoco/mujoco200, and copy the license key (mjkey.txt) to ~/.mujoco/mjkey.txt. Note that unzip of the MuJoCo binaries will generate mujoco200_linux. You need to rename the directory and place it at ~/.mujoco/mujoco200.
4. You may need to update bashrc by adding the following lines and source it:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<path/to/.mujoco>/mujoco200/bin/

Baselines repo

Clone the baselines repo in another folder outside graff/. Baselines repo: https://github.com/priyankamandikal/baselines_graff

git clone [email protected]:priyankamandikal/baselines_graff.git

Conda environment

1. Create a conda environment called 'graff':

conda create -n graff python=3.10
conda activate graff

2. Install packages:

conda install -c menpo glew 
conda install pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia
conda install tensorflow
conda install -c conda-forge matplotlib
conda install -c conda-forge moviepy
conda install -c conda-forge opencv
conda install -c conda-forge seaborn
conda install -c conda-forge quaternion
pip install open3d trimesh numba pyquaternion nvidia-ml-py3
conda install -c anaconda scikit-image

Note: It is important to install opencv after moviepy so that you don’t run into ffmpeg issues

3. Install mujocopy

conda install patchelf
pip install mujoco_py==2.0.2.13

If you face cython errors during mujocopy installation, you might want to try the following:

pip install "cython<3"
pip install lockfile glfw
pip install mujoco_py==2.0.2.5

If using GPU, after successful installation, do this: Open file <path-to-conda-env>/graff/lib/python3.10/site-packages/mujoco_py/builder.py. In function load_cython_ext(), change line 74: From:

else: 
    builder = LinuxCPUExtensionBuilder

To:

else: 
    builder = LinuxGPUExtensionBuilder 

4. Install baselines:

cd <path-to-baselines-repo>/baselines 
pip install -e . 

Verify installation

Make sure that you are able to import the following packages in python without any errors

import torch
import mujoco_py
import baselines

Dataset

We train and validate our grasping policy on 16 objects from the ContactDB dataset. We port the object meshes into Mujoco and convexify them using VHACD. The convexified meshes are present in envs/resources/meshes/contactdb/vhacd.

Training

Visual Affordance

• Render ContactDB meshes to generate images for training and testing the affordance model

    python affordance-pred/render_contactdb_data.py --obj apple cup cell_phone door_knob flashlight hammer knife light_bulb mouse mug pan scissors stapler teapot toothbrush toothpaste

• Train the visual affordance model on ContactDB

    python affordance-pred/train.py

Dexterous Grasping

• To train the graff model, run:

bash scripts/train/graff.sh

Note that this code trains on the 3D affordance points directly extracted from ContactDB. To use the visual affordances, render the predicted 2d affordance map into 3D using the depth map, save the 20 maximally diverse keypoints on the object and train the model using the predicted points.

• To train the no prior baseline, run:

bash scripts/train/noprior.sh

• To train the center of mass baseline, run:

bash scripts/train/com.sh

Evaluation

Follow the steps detailed above to set up the environment and download the pre-trained models.

Visual Affordance

• Evaluate the affordance model on rendered ContactDB objects

    python affordance-pred/evaluate_contactdb.py

Dexterous Grasping

• For evaluating the trained grasping policy, run:

bash scripts/eval.sh

The results (metrics and videos) will be saved inside expts/

Sample Results

Visual Affordance

Below are a few sample results from our visual affordance model

Dexterous Grasping

Below are a few sample results from our dexterous grasping policy.

Acknowledgements

The PPO algorithm has been adapted from https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail