This repository contains an implementation of the Sparse, Online, Locally Adaptive Regression using Gaussian Processes (SOLAR-GP) algorithm that is being presented in IEEE-RAL 2020. The code in this repository is considered research and "experimental", so results may vary depending on hardware and configuration

2D Manipulators	3D Manipulators

This package depends on the below packages in order to compile and operate:

• teleop_utils (https://github.com/bpwilcox/teleop_utils)
• baxter_interface (https://github.com/RethinkRobotics/baxter_interface)
  • May require additional baxter dependencies

Though the core SOLAR_GP algorithm does not depend on Baxter, currently the data buffer uses the EndpointState custom msg from the baxter_core_msgs package (a dependency of baxter_interface). This is may change in the future.

This package is currently based around the Rethink Robotics Baxter robot implementation of the SOLAR_GP algorithm. In the Baxter launch and Baxter scripts directories are Baxter-specific implementations of core classes.

• launch Baxter simulator in Gazebo or real Baxter robot (see )
• set teleop_device in run_baxter.launch to "xbox"
• run roslaunch SOLAR_GP_ROS run_baxter.launch

• launch Baxter simulator in Gazebo or real Baxter robot (see )
• set teleop_device in run_baxter.launch to "bag"
• set bagfile to path of trajectory rosbag
• run roslaunch SOLAR_GP_ROS run_baxter.launch

In order to the SOLAR_GP algorithm with your own robot, you need to implement derived classes of the following base classes:

• SolarTrainer
• SolarPredictor
• RobotController

In these derived classes, you can set up robot-specific variables and API, and implement robot-specific overrides of the base class functions in order to perform training, prediction, and control.

After creating the derived classes, you can create a main to initialize and run the derived classes. See the below Baxter implementations for examples:

• BaxterTrainer
• BaxterPredictor
• BaxterController