Retrieve Base Feedback (Pose, Velocity, Wrench) While in LOW_LEVEL_SERVOING Mode

[Reichenbar]

Hi Kinova team,

We have successfully implemented low-level position real-time control for the Kinova Gen3 robot arm. Our control pipeline is as follows:

1. A high-level controller runs at 10 Hz and continuously updates the desired end-effector pose.
2. Using base.ComputeInverseKinematics(input_IkData), we compute the corresponding desired joint angles.
3. These joint angles are sent to a low-level controller running at 1 kHz.
4. Before calling base_cyclic.Refresh(command), we perform necessary interpolation of joint angles to ensure safe actuator execution.

In our experiments, this pipeline works well, and the arm follows the commanded trajectories reliably. We would like to ask two questions:

1. Do you have any comments or suggestions regarding our control pipeline?
   Is this method consistent with how you recommend using LOW_LEVEL_SERVOING mode? Are there better or more stable alternatives for low-level position control?

2. Most importantly, we are unable to retrieve base feedback while in LOW_LEVEL_SERVOING mode.
   When our high-level controller (running at 10 Hz) attempts to call base_cyclic.RefreshFeedback(), it does not update critical state information, such as the end-effector pose, velocity, external force/torque (wrench), etc. These values, normally available in BaseFeedback, are essential for our robot learning algorithm. Is there any way to retrieve these feedback values?

Thanks for your support and looking forward to your advice!

[martinleroux]

Hi @Reichenbar ,

A few things.

1. The ComputeInverseKinematics() function is designed for planning. The fact that you had the foresight to run it at 10Hz instead of 1 kHz is great, but in some conditions it is possible you wouldn't even get the function to return in 100ms (those are rare, but possible). If you aim for reliable performance, you may consider using 3rd party kinematics library such as RL in C++ or Peter Corke's Robotics library for Python.

2. Note that the robot can only process one UDP call (Refresh/RefreshFeedback) per 1ms. Since the robot expects a command rate of 1 kHz, make sure you always use the feedback from the Refresh() function and do not need to call RefreshFeedback() during operation.

3. The point of low-level control is to bypass all of the controller's oversight and forward your commands directly to the actuators. The fact that the base feedback is no longer updating in low-level is therefore normal. If you need these values, you would need to compute them yourself - again I can suggest using a 3rd party kinematics library.

4. While your general strategy for low-level control looks absolutely correct, if your high level controller command rate is no faster than 10 Hz, you may consider using base.SendJointsSpeeds() or base.SentTwistCommand(), which supports an update rate of up to 40 Hz (shared with other high level functions like computeIK though). Using this, you can still have RefreshFeedback running via UDP in a parallel thread to gather all of the robot's feedback at 1 kHz, including the baseFeedback.