Differentiable Projective Dynamics

The work and examples in this project are based on the original work of https://github.com/mit-gfx/diff_pd. This is the code accompanying our paper "Sim-to-Real for Differentiable Projective Dynamics on Soft Robots: Meshing, Damping, and Actuation" to run the simulation experiments that are shown.

Recommended systems

Ubuntu 20.04 (18.04 also tested)
(Mini)conda 4.7.12 or higher
GCC 9.3.0 (7.5 also tested, other versions might work but we tested the codebase with these versions)

Installation

git clone --recursive git@github.com:srl-ethz/diffPD_sim2real.git
cd diffPD_sim2real
conda env create -f environment.yml
conda activate diffPD_sim2real
./install.sh

Examples

Navigate to python/[scenario] where scenario is one of the following:

realbeam_experiments

Here you'll find the examples for the Clamped Beam under External Force, run python [example_name] where example_name is one of the following:

clamped_beam_Case_A-1.py: Generates the results for the first case from TABLE 1 in our paper. Similarly, the other cases can also be run. Running these examples with the --video flag will create a video of the simulation as well, however, this will require much more time to run.
parameterOptimization_clamped_beam_Case_E.py: Optimize Young's Modulus of beam to fit real data.
numerical_damping_compensation_Case_A-1.py: Optimizes damping parameter such that the external force can dampen the dynamic behavior in simulation to match the real motion.

1segment_arm

Here you'll find the examples for the Soft Robotic Arm, run python [example_name].py where example_name is one of the following:

soft1arm.py: Run the simulation for accurate, high resolution single segment robotic arm mesh.
muscles_AC1.py: Run the optimization for the single segment robotic arm with muscle model AC1 from our paper.
muscles_AC2.py: Run the optimization for the single segment robotic arm with muscle model AC2 from our paper.
muscles_AC2_multiActuation.py: Run the optimization for the single segment robotic arm with muscle model AC2 from our paper using multiple actuations, every 5 frames (0.05s), now matching the dynamic behavior of the real pressure actuation better.
muscles_AC2_interpolation.py: Fits curve through optimized actuation-pressure points, then predicts muscle actuation for new pressure values, and runs the simulation to compute the final position error of the predicted values.

soft_fish

Here you'll find the examples for the Soft Fish Tail, run python [example_name].py where example_name is one of the following:

soft_fish_2muscles.py: Run the optimization for the soft fish tail with 2 muscles, one extending and one contracting, where the contraction is always half of the extension. We optimize a single muscle actuation for the whole motion of pressure actuation.
soft_fish_2muscles_multiActuation.py: Optimization for multiple actuations, every 20 frames (0.2s), now matching the dynamic behavior of the real pressure actuation better.