mmintlab/ndcf_envs

Sim. Environments for Neural Contact Field Development

NDCF Environments

Isaac Gym (Preview) environments for the Neural Deforming Contact Fields project.

Setup

We use two environments: one for running the simulation and another for post-processing the resulting meshes, pointclouds, wrenches etc. into data ready to train our networks.

Simulation Environment

Follow Nvidia's instructions here to install Isaac Gym to a conda environment. Other required packages can be found in sim_environment.yaml.

Processing Environment

Setup processing conda environment with proc_environment.yaml.

Running

Running the simulation

To collect a new dataset of simulated presses, use your simulation environment python to run the following:

python ncf_envs/sample_sim_presses_gen_terrain_proc.py cfg/primitives/ridge.yaml -o <out> -n <num> -e 1

This will run the simulation for <num> episodes and save the results to <out>. Switch out cfg/primitives/ridge.yaml for other configuration files found in cfg/primitives/ to run with different environments. This script does wraps Isaac Gym in a process to catch simulation failures that are tricky to detect and automatically starts from where it left off.

Post-Processing

Once the dataset has been collected, use your processing env python to run the following to generate SDF samples, pointclouds, meshes, etc. to be used during training and evaluation.

python ncf_envs/process_sim_data.py <output directory used by simulator> <path to tool .tet file used in simulation>

Default path for sponge is: assets/meshes/sponge/sponge_2/sponge_2.tet