P-HER

The code of paper "Trajectory Progress-based Prioritizing and Intrinsic Reward Mechanism for Robust Training of Robotic Manipulations" submitted to T-ASE.
Our code is developed based on OpenAI Baselines

Video of simulation and real-world experiments

video.2.mp4

Video of real-world applicaion (workpieces sorting task)

workpiece.3.mp4

Requirement(important)

• Python==3.6.13
• tensorflow==1.15.0
• numpy==1.19.5
• mujoco==2.0.0
• mujoco_py==2.0.13
• mpi4py==3.1.4
• gym==0.15.7
• panda-gym==2.0.0 forked from qgallouedec/panda-gym (https://github.com/weixiang-smart/panda-gym)

Installation

pip install -e .

Usage

1. Open the terminal in ./basedlines/her/experiment

2. Train the model with P-HER in PandaPickAndPlaceJoints-v2 by running the command

python train.py --env_name PandaPickAndPlaceJoints-v2  --prioritization motivation --ratio_o 0.75 --ratio 0.25 --seed 2 --n_epochs 100 --num_cpu 8 --logdir logs/PandaPickAndPlaceJoints-v2/test/7525/finaltest/r2 --logging True

Reference

• OpenAI Baselines
• Energy-Based Hindsight Experience Prioritization
• Diversity-based Trajectory and Goal Selection with Hindsight Experience Replay