Soft Actor-Critic based Path Planning for Multi-Arm Manipulators with Moving Obstacles using LSTM

Based on Soft Actor-Critic (SAC), a reinforcement learning algorithm, for moving obstacles, LSTM predicts future obstacle positions and plans the shortest path to the goal  point.

Path planning starts at a random start point at the beginning of the episode, sets a random goal point, and plans the path accordingly.

The experimental results are shown through the Gazebo simulation and the actual experimental environment implemented.

In Gazebo simulation, given an random start point and goal point, the optimal collision-free path is calculated using the SAC algorithm and LSTM.

Training is performed using Tensorflow 2.x and testing is performed using the trained network.

Operate two 3-DOF open manipulators using ROS in a dynamic environment with moving obstacles.

The testing of the real manipulator's operation using two real 3-DOF open manipulators (open-manipulator from robotis.com).

The real test environment used obstacles moving along the y-axis due to realistic constraints.

At the beginning of the episode, a random goal point is given and the optimal collision-free path is calculated using the SAC algorithm and LSTM.

A camera estimating the position of an obstacle was installed on the test environment, and obstacle position data was obtained using machine vision.

It is verified that the proposed method indeed generates the collision-free path for the real manipulator with both static and moving obstacles.