ros_controllers

Dependencies

• Python 3.9.10 with TensorFlow Agents version 0.10.0
• ROS Melodic
• libfranka 0.9.0
• franka_ros 0.9.0 + commits
• Tested with Franka Emika Panda Robot system version 4.2.1

Installation and Policy Execution

1. Create a ROS catkin workspace with the following packages:

• this (softgrasp_ros_controllers),
• softgrasp_ros,
• our fork of franka_ros

2. Download a checkpoint from here and unzip it in the policies directory. Denote its path as policies/$EXP_NAME. Also note the $CKPT_NAME from that table.
3. Activate the catkin workspace created above, and then launch: roslaunch softgrasp_ros_controllers policy_name:=softgrasp robot_ip:=<robot IP> exp_name:=$EXP_NAME ckpt_dirname:=$CKPT_NAME.

Notes

• src/tf_policy_controller.cpp implements a ROS controller that
  • creates a "goal" for the policy server using robot model and state
  • composes the current EEF pose with pose delta predicted by the policy server to get EEF target
  • runs high-frequency task space impedance control with the EEF target
• policy server is implemented by Python3 nodes in nodes
• policy observation and action are in (pos, quat) format
• chrony time synchronization

sim-to-real alignment

• Match kp from robosuite osc_pose.json with franka_example_controllers/cfg/compliance_param.yaml on the realtime ROS computer
• Match {input,output}_{min,max} from osc_pose.json with those variables in TFPolicyActionServer

Circle sanity check

• Run sim.launch or real.launch with policy_name:=circle
• rosbag record tf tf_static -o $HOME/circle.bag
• roslaunch softgrasp_ros_controllers circle_sanity_check.launch bag_filename:=$HOME/circle.bag

Static Poses

They are stored in launch/static_poses.launch.

• cTw contains the output of marker_pose from softgrasp_ros with just the table-affixed marker visible
• cTg contains the ouput of marker_pose from softgrasp_ros with the table-affixed marker occluded and goal marker card visible
• When the camera location is moved, update both cTw and cTg static transform broadcasters
• Procedure for wTb:
  • keep just cTw uncommented
  • roslaunch softgrasp_ros_controllers static_poses.launch
  • roslaunch realsense2_camera rs_camera.launch
  • roslaunch <camera_pose.launch>, where camera_pose.launch is the launch file generated by the hand-eye calibration
  • rosrun tf tf_echo world_table panda_link0
  • Copy-paste its output to wTb, remembering to either normalize the quaternion, or if using Euler angles, reverse the order of RPY because TF's static transform broadcaster requires YPR
• The camera_link_broadcaster node in <camera_pose.launch> contains the info generated by hand-eye calibration