Implementation of Neural-based Motion planner for Motion planning of Panda Arm robot. There are two variations of Motion planner:

• Motion planning in a single familiar environment.
• Motion planning in multiple unknown environments.

The Project report can be found here.

• Data Generation.

  • Any Classical Motion Planner can be used to generate dataset. Here we use OMPL Motion Planner to generate RRT*.
  • We use MoveIt Platform to integrate Motion planning with visualization and simulation.
  • Custom dataset class.

• MPNet algorithm.

  • Network architecture for end-to-end and modular trained networks.
  • Software stack for training and testing the algorithm.

• Python visualization files to compare performance of MPNet with the expert demonstrator.

• ROS.
• Gazebo.
• MoveIt.
• Pytorch.

• In case of motion planning in a single known environment, we only need the absolute initial and final states that completely define the configuration of the robot.
• In case of motion planning in multiple unknown environments, we consider the point cloud data along with the initial and final configurations. we train the network using 7 environments and test it's performance on the trained environments. In addition to that we test the performance on 3 new environments.

• For training to generate motion plans in a single environment.
  • Open RViz roslaunch panda_moveit_config demo.launch.
  • Run the joint state planner to generate the dataset. python panda_arm_motion_planning/scripts/joint_space_planning.py. You can tweak the parameters according to your choice.