Home

Learning "near-contact" grasping strategy with Deep Reinforcement Learning

This is an implementation of Deep Deterministic Policy Gradient from Demonstration (DDPGfD) to train a policy to perform "near-contact" grasping tasks, where object's starting position is random within graspable region. We took one "near-contact" strategy from this paper as expert demonstration and train a RL controller to handle a variety of objects with random starting position.

This environment runs on MuJoCo with an intergration of OpenAI gym to facilitate the data collection and traning process.

Table of Contents

• Getting Started
• Software Setup

• Common Issues + Fixes

• How to run the code and experiments
• Data and input variations
• Reinforcement Learning Overview

• DDPGfD Algorithm
• Policy Training
• Training Pipeline
• Controllers
• Evaluation

• Video Explanations and Tutorials