Tactile Pushing using Reinforcement Learning

Installation

This packages has only been developed and tested with Ubuntu 20.04 and python 3.9.

Installing tactile gym:

git clone https://github.com/maxyang27896/tactile_gym_push.git
cd tactile_gym_dev
pip install -e .

Installing mbrl-lib

cd tactile_gym_dev/utils/mbrl-lib
pip install -e .

Install PyTorch from the installation page.

Training Mode-Free RL Agents

Run: The pushing environment has been trained using SAC from stable-baselines3 for all training, helper scripts are provided in tactile_gym/sb3_helpers/. Inside the /sb3_helpers directory, run

python train_agent.py

Progress plots and saved models will be in /sb3_helpers/saved_models/object_push-v0 folder.

Training Mode-Based RL Agents

For model-based RL, the agent has been trained using PETs from mbrl-lib. Some examples of usage are provided in the /utils/mbrl-lib/notebooks folder.

To start a training, inside the /mbrl_helpers directory, run

python object_push_training.py --num_trials 100

The models and progress will be saved under /mbrl_helpers/training_model by default.