biorobotics/Wombat_robosuite

This work builds on top of mujoco-powered robosuite framework, and is designed for learning object-agnostic control policy for pick-and-place operations on a novel 6-DoF parallel manipulator using deep reinforcement learning. This pipeline utilizes DDPG combined with HER algorithm for training a robust manipulation policy.

🤖 Wombat_robosuite

Simulation framework for learning an object-agnostic control policy for pick-and-place operation on a novel 6-DoF parallel manipulator (aka D3 arm) using Deep reinforcement learning, specifically DDPG (Deep Deterministic Policy Gradient) + HER (Hindshight Experience Replay) algorithm.

🛠️ Installation

• Install MuJoCo

1. Download the MuJoCo version 2.1 binaries for Linux.
2. Extract the downloaded mujoco210 directory into ~/.mujoco/mujoco210.
3. Clone this repository

cd ~
git clone -b main https://github.com/biorobotics/Wombat_robosuite.git

4. Install the dependencies

cd Wombat_robosuite/robosuite
pip3 install -r requirements.txt

5. Test your installation with

python demos/demo_random_action.py

🔄 Pipeline

🏋️ Training (DDPG+HER)

cd ~/Wombat_robosuite/robosuite/HER+DDPG/hindsight-experience-replay-latest
python D3_train_dyn_rand.py

Weights will be stored in 'saved_models/PickPlaceiPhone'. Currently, the best trained weights are with the name 'Model_dyn_rand_epoch_50.pt'.

🎮 Demo (Testing)

cd ~/Wombat_robosuite/robosuite/HER+DDPG/hindsight-experience-replay-latest
python D3_demo_dyn_rand.py

📊 Results

🙏 Credits

This project builds upon the excellent work done in the Robosuite Framework. Many thanks to all the contributors involved.