This repository contains the implementation for the paper An End-to-End Differentiable Framework for Contact-Aware Robot Design (RSS 2021).

In this paper, we propose a fully differentiable pipeline to jointly optimize the morphology and control of manipulator robots. At the core of the framework is a deformation-based morphology parameterization and a differentiable simulation.

The framework itself is general and not limited to manipulator robots, we select the case study of manipulator robots because of its complexity and contact-rich nature. Welcome to try our code on any other types robots as well.


We provides two methods for installation of the code. Install on local machine and Install by Docker.

Option 1: Install on Local Machine

Operating System: tested on Ubuntu 16.04 and Ubuntu 18.04

  1. Clone the project from github: git clone https://github.com/eanswer/DiffHand.git --recursive.

  2. Install CMake >= 3.1.0: official instruction for cmake installation

  3. build conda environment and install simulation

    cd DiffHand
    conda env create -f environment.yml
    conda activate diffhand
    cd core
    python setup.py install
  4. Test the installation

    cd examples
    python test_redmax.py

    If you see a simulation rendering with a two-link pendulum as below, you have successfully installed the code base.


Option 2: Install by Docker

We provide a docker installation in the docker folder. Follow the readme instruction in docker folder to complete the installation.

Code Structure

There are two main components of the code base:

  • Differentiable RedMax: DiffHand/core. The differentiable redmax is based off RedMax and further makes if fully differentiable. It provides the simulation derivatives w.r.t. both simulation parameters (kinematics- and dynamics-related parameter) and control actions. It is implemented in C++ for computing efficiency. We provide a simulation document for mathematical details of our differentiable RedMax.
  • Morphology and Control Co-Optimization: DiffHand/examples. We build an end-to-end differentiable framework to co-optimize both the morphology and control of manipulators. We use L-BFGS-B as our default gradient-based optimizer and also provides the source code for the gradient-free baseline methods.

Run the Code

It is recommended to try out the scripts in play with redmax simulation first if you would like to get familiar with simulation interface.

Run the examples in the paper

We include the four co-design tasks from the paper in the examples folder.

  • Finger Reach
  • Rotate Cube
  • Flip Box
  • Assemble

To run the L-BFGS-B optimization with our deformation-based design parameterization, you can enter the corresponding folder and run demo.sh under the folder. For example, to run Finger Reach,

cd examples/rss_finger_reach
bash demo.sh

Run batch experiments of baseline algorithms

We include the gradient-free baselines (except RL) and the control-only baseline in this repository. For the RL baseline, we use the released code from Luck et al with some modifications to our proposed morphology parameterization.

To run the baseline algorithms or our method in a batch mode, enter the corresponding folder and run run_batch_experiments.py. For example, to run Flip Cube with CMA-ES,

cd examples/rss_finger_flip
python run_batch_experiments.py --method CMA --num-seeds 5 --num-processes 5 --save-dir ./results/

Play with redmax simulation

We provide several examples to test the forward simulation and its differentiability.

  • examples/test_redmax.py provides the script to show how to run forward simulation and rendering. It can be easily executed by:

    python test_redmax.py --model hand_sphere

    Here, you can also try other models provided in assets folder (models are described by xml configuration files).

  • examples/test_finger_flick_optimize.py provides an example for using the backward gradients of the simulation. In this example, we use gradient-based optimization to optimize the control sequence of a pendulum finger model to flick a cube to a target location. run it by:

    python test_finger_flick_optimize.py

    The initial control sequence is shown first and you can press [Esc] to close the rendering and start the optimization. After successful optimization, you will see a rendering as below:



If you find our paper or code is useful, please consider citing:

    AUTHOR    = {Jie Xu AND Tao Chen AND Lara Zlokapa AND Michael Foshey AND Wojciech Matusik AND Shinjiro Sueda AND Pulkit Agrawal}, 
    TITLE     = {{An End-to-End Differentiable Framework for Contact-Aware Robot Design}}, 
    BOOKTITLE = {Proceedings of Robotics: Science and Systems}, 
    YEAR      = {2021}, 
    ADDRESS   = {Virtual}, 
    MONTH     = {July}, 
    DOI       = {10.15607/RSS.2021.XVII.008}