README.md

Pre-requisites

1. Install MuJoCo as explained in here
2. Install mujoco-py as explained in here or mujoco-py github
3. Tensorflow 2.x to run DDPG

Installation

1. Clone the repository and inside spaceRobot_RL folder (where the setup.py is), run pip install -e .

Test Installation

import gym
import spacecraftRobot
env = gym.make('SpaceRobot-v0')

Test DDPG

python train_spaceRobot.py

Test POLO

Plan Online Learn Offline is a value function learning paper using trajectory optimization. Details availabe at POLO

1. cd POLO
2. python job_script1.py

Note: Needs mujoco-py

The debris could be stationary or rotating about some arbitrary axis. The aim is to capture the debris and bring it back to a station.

The big cube is assumed to be the spacecraft and carries a robot arm which has seven rotary joints and has a gripper with two linear joints.

Reward:

Currently it is composed of three parts.

dist = distance between end-effector and debris # to reach near the debris

act = 0.0001 * np.dot(action, action) # to obtain minimal actuator torques

rw_vel = 0.001 * sum of squares of base linear and angular velocites # for reaction-less control

reward = -(dist + act + rw_vel)

Note: As of now there is no reward for grasping

Current issues:

1. How to set reward for grasping? 
Is it could be just negative of the distance between the end-effector and debris 
or add a camera and get images at every instant of time?

Ideas

1. Demonstrated replay buffer
2. Reaction-less control