PANDA GYM INTEGRATION

This package connects panda_gym with the real robot (Panda Franka Emika) through ROS

Table of Contents

• Getting started
  • Dependencies and building
• Files
  • panda_gym_run
  • ai_controller
  • panda_path
  • other
• VSCode
• Author
• License

---

Getting started

This package was tested with real robot with:

• Ubuntu 20.04 LTS Focal Fossa
• ROS noetic
• libfranka 0.8.0
• franka_ros 0.7.1
• panda_moveit_config 0.7.5
• panda_controller noetic_dev
• frankx master on cc81e52045be280ee2f6b928e954d05ccc7798ad

Dependencies and building

Follow Dependencies and building section here

frankx

Install frankx

Note: To make frankx compatible (which assumes libfranka 0.7.1) with libfranka 0.8.0, go to frankx/src/frankx/robot.cpp and change setEE in the setDefaultBehavior() method with:

setEE({1.0, 0.0, 0.0, 0.0,  0.0, -1.0, 0.0, 0.0,  0.0, 0.0, -1.0, 0.0,  0.0, 0.0, 0.0, 1.0});

Remember to always call setDefaultBehavior() whenever you use frankx. Example:

robot = Robot("192.168.1.2")
gripper = robot.get_gripper()
robot.set_default_behavior()
robot.recover_from_errors()

panda_gym:

The AI has been trained with the goal closest to the default one...

Go to panda_gym/envs/assets/pickandplace.json and set the object -> body -> basePosition to [0.025, 0.025, 0.4] for better results.

Get panda_gym_integration:

cd ~/panda_ws/src/

pip3 install --user numpy panda_gym torch

git clone https://github.com/Envq/panda_gym_integration.git

cd ~/panda_ws

catkin build

---

Files:

panda_gym_run

• This file allows you to run panda_gym and test your policy defined in src/panda_actors.py

• Parameters available:

  • DEBUG_ENV_ENABLED: enable panda_gym environment debug prints
  • DEBUG_AI_ENABLED: enable ai actor debug prints
  • NUM_EPISODES: number of episodes to run
  • LEN_EPISODE: maximum number of steps per episode
  • WRITE_ENABLE: if true, it asks you whether to save the path generated by the episode performed
  • FILE_NAME: specify here the name of the file containing the path to be saved
  • ACTOR: specify here your actor

ai_controller

• This file allows you to use your AI to control the real robot with frankx

• Parameters available:

  • IP: specify here the ip of the robot
  • DEBUG_ENV_ENABLED: enable panda_gym environment debug prints
  • DEBUG_AI_ENABLED: enable ai actor debug prints
  • NUM_EPISODES: number of episodes to run
  • LEN_EPISODE: maximum number of steps per episode
  • DYNAMIC_REL: if this value is 0.08, it set velocity, acceleration and jerk to 8% of the maximum
  • ACTOR: specify here your actor
  • START_POSE: this is a list of start pose for your episodes. Note**: the AI used requires that the start pose is always the same
  • OBJ_POSE: this is a list of object pose for your episodes
  • GOAL_POSE: this is a list of goal pose for your episodes

• Note: the accuracy of the AI is in the order of 10 cm.

panda_path

• panda_path_frankx (or panda_path_moveit) is a file that generates a trajectory with the path contained in the specified file and executes it with frankx (or moveit)

• Parameters available for panda_path_frankx:

  • IP: specify here the ip of the robot
  • DYNAMIC_REL: if this value is 0.08, it set velocity, acceleration and jerk to 8% of the maximum
  • FILE_NAME: is the name of the file in data/paths where read the path
  • TIME_DELAY: is the delay after which to add new waypoints in "async" mode

  execute:

  • python3 panda_path_frankx.py planning ---> to generate the trajectory offline
  • python3 panda_path_frankx.py linear --------> to move to each point at runtime with linear movements
  • python3 panda_path_frankx.py planning ---> to generate the trajectory online receiving waypoints with a TIME_DELAY delay

• Parameters available for panda_path_moveit in launch/panda_path.launch** (Note**: execute this file with roslaunch panda_gym_integration panda_path.launch):

  • arm_speed: is speed used for setMaxVelocityScalingFactor() method of Moveit

  • gripper_speed: is the speed used for the move() method of the gripper

  • gripper_speed, grasp_speed, grasp_force, grasp_epsilon_inner, grasp_epsilon_outer: are the values used in grasp() method of the gripper

  • eef_step, jump_threashould: are the values used in compute_cartesian_path() method of Moveit

  • real_robot: if true, panda_controller will use franka_gripper's action clients

  • file_name: is the name of the file in data/paths where read the path

  • mode = 1: the generation of the trajectories is done with the compute_cartesian_path() method of Moveit

  • mode = 2: each point is read and immediately executed with moveArmPoseTCP() method of panda_controller

  • mode2_delay, mode2_delay: If mode is 2, each point is execute with mode2_delay and with go(wait=mode2_wait) in Moveit

other

• debug_frankx: use python3 debug_frankx.py cmd_name, where cmd_name is the command name you want to test with frankx
• src/panda_actors: this files containes the policies and AI, used in panda_gym_run and ai_controller
• ai/: this folder contains the necessary files for the AI taken [here] (https://github.com/LM095/Multi-Subtask-DRL-for-Pick-and-Place-Task)

---

VSCode

I used these extensions:

• c/c++ by microsoft
• c++ intellisense by austin
• cmake by twxs
• doxygen by christoph schlosser
• clang-format by xaver
• doxgen documentation by christoph schlosser
• python by microsoft
• git graph by mhutchie
• gruvbox mirror by adamsome
• vscode-icons by icons for visual studio code
• git graph by mhutchie

---

Author

Enrico Sgarbanti @Envq for panda-gym integration with real robot

Luca Marzari @LM095 for AI. look here for more details

License

This project is licensed under the GPL v3 License - see the LICENSE.md file for details