This directory contains network architecture definitions for the Grasping critic architecture described in QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation.
The following command trains the QT-Opt critic architecture for a few gradient steps with mock data (real data is not included in this repo). The learning obective resembles supervised learning, since Bellman targets in QT-Opt are computed in a separate process (not open-sourced).
git clone https://github.com/google/tensor2robot
# Optional: Create a virtualenv
python3 -m venv ~/venv
source ~/venv/bin/activate
pip install -r tensor2robot/requirements.txt
python -m tensor2robot.research.qtopt.t2r_models_test
This directory also contains a multi-task optimization method PCGrad that is implemented in the form of a optimization wrapper. This is based on the open-source implementation here.