Tune the actuator parameters for the mujoco model

giotherobot commented 8 months ago

For now all the actuators in the model have common parameters (kp=100).

It would be better to tune these.

giotherobot commented 7 months ago

With Kp=1000 ergoCub is able to stay upright:

giotherobot commented 7 months ago

Just to write it down otherwise I forget it, a basic sim2real step (that probably we can do in a second step) is to align the Kp used for position control with the one used on the real robot. Disregarding friction effects, it should boils down to get the gains used on the real robot, and convert their units (accounting for the motor K_t current/torque gain and transmission ration) to the one used in MuJoCo . Francesco Romano (a former member of the lab) did that some time ago for Gazebo, but I can't remember where that code is.

https://github.com/ami-iit/element_llm-task-planner/issues/34#issuecomment-1983422432

For reference @traversaro

traversaro commented 7 months ago

Just to write it down otherwise I forget it, a basic sim2real step (that probably we can do in a second step) is to align the Kp used for position control with the one used on the real robot. Disregarding friction effects, it should boils down to get the gains used on the real robot, and convert their units (accounting for the motor K_t current/torque gain and transmission ration) to the one used in MuJoCo . Francesco Romano (a former member of the lab) did that some time ago for Gazebo, but I can't remember where that code is.

#34 (comment)

For reference @traversaro

I can't find that code unfortunately.

giotherobot commented 7 months ago

Regardless of the units conversion, where can I find the gains of the gazebo controllers or the oens used in the robot?

traversaro commented 7 months ago

Sure! Let's use the ergoCub l_knee as an example. In both Gazebo Classic and Modern GAzebo, its gains are contained in https://github.com/icub-tech-iit/ergocub-software/blob/2b3dff9b264a7fac89e9608d5e4e40dd24fcba21/urdf/ergoCub/conf/gazebo_ergocub_left_leg.ini#L20. The unit of those values are Nm/[deg] for Kp, and similar units (but with a s somewhere) for Ki and Kd, see https://github.com/robotology/gz-sim-yarp-plugins/blob/1c38e89a3303d987bae80e3b58c296b14105ee86/plugins/controlboard/src/ControlBoard.cpp#L447 and https://github.com/robotology/gz-sim-yarp-plugins/blob/1c38e89a3303d987bae80e3b58c296b14105ee86/plugins/controlboard/src/ControlBoard.cpp#L600-L604 for more details or ask to @xela-95 .

For what regards the real robot, it is a bit more complex. First of all, there are different ways for controlling in position the robot, depending on the output of the position controller, if the output of the position controller is current (that is then sent to a low level current control loop) or PWM (that is directly sent to the motor). The gains for the ergoCubSN001 motor (the one used in IIT robot arena as of April 2024). From https://github.com/ami-iit/robots-configuration/blob/9849c5617ab1528a92ba1193e47496a403aa2724/ergoCubSN001/hardware/mechanicals/left_leg-eb8-j0_3-mec.xml#L9 you get that l_knee is the fourth axis (i.e. with index 3) of the eb8-j0_3 board. From https://github.com/ami-iit/robots-configuration/blob/9849c5617ab1528a92ba1193e47496a403aa2724/ergoCubSN001/hardware/motorControl/left_leg-eb8-j0_3-mc.xml#L46-L57 it seems that the position loop uses pwm as output, but I am not sure about this. However, assuming that pwm is indeed the right one, we need to understand:

Unit of PWM: I guess this is tipically arbitrary, the important is that the corresponding k_tau is using the same unit
Input position: the code reports metric_units, but I guess we need to double check this.

So, assuming that k_p units is [PWMUnit]/[deg], probably we need to find a k_tau (and a transmission ratio, depending on how k_tau is defined) for which we need to multiply the gains. I guess we can either ask to @isorrentino and @GiulioRomualdi, or directly ask a question in https://github.com/robotology/community where people from iCub Tech can reply, thanks!

GiulioRomualdi commented 7 months ago

A possible starting point could be https://github.com/ami-iit/element_sensorless-torque-control/issues/114

traversaro commented 7 months ago

A possible starting point could be ami-iit/element_sensorless-torque-control#114

Thanks! @isorrentino what is the unit of the ktau in https://github.com/ami-iit/element_sensorless-torque-control/issues/114#issuecomment-1948247808 ? Nm/A, Nm/PWMUnit or something else? Are those Ktau also accounting for the transmission ratio?

S-Dafarra commented 7 months ago

Back in time we started an effort to convert the gains from pwm to metric. I can't seem to find the issue at the moment, but I remember using an Excel file to convert the gains, so there should be a fixed conversion rule. The problem back in time was that the converted gains worked terribly (we thought about an accuracy issue at the FW level since in metric the numbers were much lower), so we never switched to metric.

S-Dafarra commented 7 months ago

I can't seem to find the issue at the moment

Found it! https://github.com/ami-iit/component_wholebody-teleoperation/issues/57 (again the importance of documenting things)

giotherobot commented 6 months ago

The tuning of the actuator is important for the stability of the simulations.

giotherobot commented 4 months ago

This can be moved to https://github.com/ami-iit/mujoco-urdf-loader

ami-iit / mujoco-urdf-loader

Tune the actuator parameters for the mujoco model #5