This repository hosts all the models underlying the iCub Firmware.
- ๐ Browse our online documentation. ๐ง Work in progress ๐ง
This repository is maintained by:
![]() |
@pattacini |
We accept contributions via forks and pull-requests.
Models underlying the iCub Firmware
Home Page: https://robotology.github.io/icub-firmware-models/
License: BSD 3-Clause "New" or "Revised" License
This repository hosts all the models underlying the iCub Firmware.
This repository is maintained by:
![]() |
@pattacini |
We accept contributions via forks and pull-requests.
Probably, this is not a good candidate task to be implemented in Simulink as therein we assume that we already receive the elaborated position.
Nonetheless, we should integrate somehow (maybe in the hand-coded FW) what done by @ale-git in https://github.com/robotology/icub-firmware/blob/master/emBODY/eBcode/arch-dspic/board/2foc/appl/2FOC-V3/src/qep.c#L150-L198.
Following up on #45, we're required to update the control model to deal with:
๐๐ป๐ Once done, we need to update https://icub-tech-iit.github.io/documentation/wrists/wrist_mk2/?h=wrist.
cc @ale-git @salvi-mattia @valegagge @fiorisi @mfussi66 @Nicogene @marcoaccame
Follow-up of #54.
From @valegagge:
I'd like to suggest improving the code inside the application in order to automatize as much as possible this check of performance: for example, we could have a simple Matlab script that could get the data collected in the firmware and plot them automatically. ๐๏ธ
At the moment there are two projects that involve the AMCBLDC board, the official one and the Experimental one.
The Experimental project involves a major refactor of the supervisor and leverages the Common
Simulink project to make the structure more modular.
While the first trials were successful (#84), the integration tests were not exhaustive.
This issue aims to track which tests need to be performed, before making Experimental the official project.
While working on issue #40, I noticed that the SupervisorTX supervisorTX doesn't handle the overcurrent flag, while the SupervisorRX performs the check.
SupervisorTX | SupervisorRX |
---|---|
![]() |
![]() |
As you can see, in the dictionary, Flafs struct doesn't contain the overcurrentinfo | The Fault handler state machine checks the filtered current value |
I suppose this is a bug, isn't it? @pattacini @mfussi66
I can fix it before addressing the #40.
When #53 is merged, we need to make the dictionary general enough to contain structures dealing w/ 2 (or even more) motors.
Get accustomed to all the details of the supervisor.
As per the title, we aim to check whether there are limitations in running 2 MBSE FW inside the CM4 as we need to control 2 PMSM motors.
It's important to update the CAN protocol so that it'll contain all the missing packets.
Also, we would need to link to it from this website's documentation.
While working on AMCFOC Testing I tried to run a motor control simulation, using the latest devel
branch.
To my surprise, the simulated pmsm could not spin because there was no commutation between phases:
We can see that 3-phase currents and voltages remain somewhat constant, while they actually should have this kind of waveform:
So I checked the AMCBLDC project, to verify if upstream everything was working fine.
Most specifically I used the testing model https://github.com/robotology/icub-firmware-models/blob/devel/boards/amcbldc/tests/control/control_pmsm.slx , and I noticed the same behaviour happening starting from the PR regarding D-axis alignment, even though the physical motor spins with no problem.
I can make some tests to see which changes in this PR affect the simulation.
We recently refactored the motor control supervisor (see #65). We now have to test the new architecture on the AMC-BLDC board.
For this we need:
Demo with setup running with the new firmware.
Originally opened by @fiorisi in different repo.
As suggested by @pattacini, it will be useful to create the AMCBLDC customizations in a separate path, so that the two versions can live side by side.
This is the main collector for issues on the development of the architectural model for the AMCFOC board.
We will mainly inherit the model from the AMCBLDC and adapt it to a different context.
Some preliminary work on this includes:
Note
๐๏ธ Regarding the timeline, we ought to complete the development by the end of PI16 (~ Oct 2023) to be ready for testing the new ergoCubv2.0 arm at the end of the year.
We're required to verify that the direct current
To this aim, we may rely on a temporary hack for the CAN protocol.
To be checked on the setup.
(Aside from
cc @Nicogene
Following up on the excellent work done on the dictionary (see #54), we need to deal with the motor-independent part of the Supervisor.
Important
It's true that we aim to address the AMCFOC, but it's essential to keep in mind that at a certain time we'll have also to integrate the AMCBLDC CAN handler and 4 motors (AMC4DC) instead of 2.
As of now, Vcc
and Vmax
are statically defined in the model. This prevents us from adapting those parameters to the setups/robots w/o re-generating the code.
It'd be thus great to have the possibility to tune the voltage values on the fly via CAN, for example to adapt for cable or battery voltage supply.
We haven't done that yet as the current CAN protocol doesn't foresee this option.
We may consider at this stage including other motor-related useful parameters to configure via CAN; see also #14.
In total:
swapBC
, hall sensor offset
.As of now, we only send the quadrature voltage and a flag reporting the status of the external e-stop.
However, the STATUS
message has many more flags that are meant to be transmitted, although many are not yet handled internally.
In view of AMC-FOC, we need to extend Supervisor_RX to deal w/ 2 motors.
To this end, it would be nice to have some proposals to discuss highlighting the major parts needing overhaul.
Indeed, it seems that it happens also with the upstream architectural model, see the video.
Given that this workflow is very unlikely to happen, it might be useful to notify it in a separate issue with low priority.
Originally posted by @mfussi66 in #84 (comment)
Following the work done here robotology/icub-firmware#360 (comment), we want to update the supervisor in order to manage the motor-hal errors: hardware-overcurrent
and hall-sequence-error.
Dod
Note: the test is not simple because this error are hw-related, so we need to find a way to simulate them.
It's be super useful to have the AMCBLDC board capable of reading the supply voltage Vcc
on the fly in order to adapt wrt. changing conditions (network supply, batter, setup...).
To this end, we're required to develop a dedicated API and expose the value to the architectural model.
See also #8.
Following the creation of the wrist controller and its integration in the ems fw, we want to publish and document the model and its generated fw for coupling/decoupling of eCub wrist.
Please see also this integration activity.
Move the controllers and the estimators to a higher-level project.
The idea of the AMCFOC motion controller is to be referenced in a For Each subsystem. This kind of subsystem allows calling the block multiple times, depending on the dimensionality of the input/output signals. In this way, we can use the same functions for multiple motors.
However, this type of subsystem relies on the internal function-call subsystems to have the trigger signal of width=1. The motor-specific supervisor is full of these triggers. It is therefore necessary to replace this kind of logic with an equivalent one, most likely with a less modular state machine, that allows us to resolve this limitation and enable the full features of the supervisor.
After this, it will be possible to create unit tests for the motor-specific supervisor.
The FOCInnerLoop block is the one called by the ADC interrupt in the firmware. Since the for-each subsystem expect to call this function once for each motor every ~36usec, we need to check if there are any implementation mismatches between the codegen and the expected behaviour which is interrupt based and not periodic.
From @marcoaccame
Another thought:
I am a little nervous to have an important high priority and high frequency IRQ Handler that executes for long time. The rule of thumb is: the more important the IRQ handler, the less code it must run. Now: it is executed every 37 us, lasts about 15 us. Also it runs code which requires mutual exclusion. A little too much in my opinion.
I would clean up and move out of the IRQ Handler all the MBD code by doing the following.
- keep inside the IRQ Handler only a quick copy from DMA area of the three currents (maybe into a circular FIFO) and use an RTOS event to wake up a very high priority thread called threadFOC.
- the high priority thread threadFOC is waken up by an event from the IRQ handler which: copies just the three currents in a thread safe and very quick mode (I may do it perhaps also w/out disabling the IRQ...) and then use the three currents to run the MBD FOC algorithm.
- redefine the mutex used by the MBD code to be an RTOS mutex because it must protect data between two RTOS threads. The advantage is that the RTOS mutex has priority inversion.
- re-evaluate the duration and frequency of the IRQ Handler and of the two threads FOC and TICK.
- if the IRQ Handler is still not regular, I would check the priorities of the IRQ handlers and of the DMA to make sure no other has higher priority.
The framework of the application is a little messy. It was developed long ago and it needs maybe some tidy up.
For now I document in here what we need to have.
Figure. Scheme of the required SW entities.We need a new thread
tFOC
alongside to the existing threadstCAN
andtTICK
. Also we need a different mutual exclusion mechanism between actions performed insidetTICK
andtFOC
: we can use aembot::os::rtos::mutex_t
.The interactions amongst entities are the following:
tCAN
->object
: when we have a new can frame from theembot::hw::can
driver then threadtCAN
, the producer, adds it to the input FIFO which is protected w/ a mutex vs concurrent access fromtTICK
, the consumer.tTICK
->tCAN
: if there is a can frame to send back as a reply or a periodic streaming the threadtTICK
sends an event to threadtCAN
which retrieves a vector of replies from the output FIFO (mutex protected).tTICK
<->tFOC
: access to shared data is done by means of artos::mutex_t
.IRQ handler
->tFOC
: the handler every 37 us has three new current values, it copies them into a volatile variable and then sends a wake up rtos event to threadtFOC
. The thread wakes up, disable IRQ for the handler, copies the volatile variable into local memory and re-enables IRQ. At this time it runs safely the FOC algorithm.The
tTICK
thread is periodic @ 1ms, thetFOC
is activated by the periodic IRQ handler @ 37us (so the thread is de-facto periodic as well).Almost all objects are available. I just need to refactor and test.
In the AMCBLDC firmware there isn't an initial calibration procedure to find the index of the optical encoder. A similar procedure is already available on the 2FOC firmware.
The procedure could be implemented in the model, not in the HAL layer.
#64 introduced the improvements detailed in #64 (comment).
We ought to apply them to AMCFOC too.
@mfussi66, I can point you to the relevant options.
cc @Nicogene
Following up on #14, we're required to:
So far, we have just published the webview of the architectural model at https://robotology.github.io/icub-firmware-models.
In #54 (comment), @mfussi66 found out that the For-Each approach is limited and cannot be applied straight to our case.
It'd be advantageous to poke around to see whether we can resort to methods better than our present solution based on block duplication.
Setting the FOC Ts to 45us (instead of 37.5us , ndr) makes the motor spin with more oscillations, is this expected?
Originally posted by @mfussi66 in #84 (comment)
Following up on #60 (comment), we're required to:
A F2F chat with @pattacini and @maggia80 highlighted that it is important to consider how the PWM signal is generated (whether it is in Left-aligned mode or Center-aligned), and how the current is measured.
Since we are dealing with an inductive load, we need to consider exponential transitory effect, during which the PWM signal of a phase is positive.
The current is measured in the middle of the positive PWM, during which the transitory might not be exhausted. This depends on the time constant (Phase resistance and inductance), and on the duty cycle.
Therefore accounting for the electrical transitory period might be the key to improve the model and reduce the performance discrepancy in continuous current (aka Iq in the case above).
Support material:
Originally posted by @mfussi66 in #36 (comment)
Let's proceed with identifying the parameters $R_t$ and $\tau$ robustly by covering a wide range of operating conditions with currents and temperatures (it'll certainly take a while to do so).
As usual, the validation will need to be performed on motors/setup/operating-conditions not seen during fitting.
Originally posted by @pattacini in #34 (comment)
It could be beneficial to rethink the mutexing used in the generated code by resorting to Service I/F.
Watch https://www.youtube.com/watch?v=Y-6JvmTq21w.
Simulink Test
is a blockset specifically devoted to testing.
Since we have already developed a bunch of nice tools for communicating with the AMCBLDC board, Simulink Test will seamlessly integrate with them to carry out preliminary regression tests like:
Also, we can produce nice reporting.
There's no hurry but we should definitely address it.
Here below some nice webinars we can inspect illustrating the perks of using a test harness and HIL:
We aim to extract general-purpose components from the AMCBLDC project and put them in a shared basket/library to support other boards' FW.
Warning
We shall guarantee that AMCBLDC will remain 100% operational throughout the process with continuous testing.
Now that we have achieved some sort of expertise regarding code generation for an embedded board, it might be nice to create a .md document that highlights the best practices and settings to apply for an effective and flexible process.
For example, it could contain:
While I was looking at the generated code of the FOC in the AMCBLDC, I spotted that the trigonometric functions for sin
and cos
are implemented by resorting to std
instead of CMSIS.
I spent quite some time investigating this issue and, from the code generation report, it seems that the replacement library for the Cortex-M correctly detects the presence of sin
and cos
although it doesn't apply the substitution.
To check the performance gain, we can simply replace std::sin
and std::cos
with arm_sin_f32
and arm_cos_f32
. However, the problem still remains when it comes to code generation.
There are easy alternatives that recruit proper approximations of sin
and cos
but the reason why the link to CMSIS is not performed is unclear.
Definitely, worth investigating.
Let me try to recap what I got from a F2F alignment with @sgiraz.
cc @Nicogene @sgiraz @marcoaccame @maggia80 @mfussi66
Originally posted by @pattacini in #67 (comment)
We're required to update the Simscape model of wrist-mk2 by exporting the latest mechanics from CREO.
cc @ale-git @salvi-mattia @valegagge @fiorisi @mfussi66 @Nicogene
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.