June 1, 2019

Mini Cheetah at ICRA

I was just at ICRA 2019 with the robot, to present a paper and demo the robot:

Not my picture.  Pulled from google images.

Here are a few other peoples' videos from the conference and the legged robot workshop.  Despite being half the size of all the other robots, Mini Cheetah was running circles around them.

At one point, we had Mini Cheetah, Spot MiniANYmalLaikagoVision 60, and Salto all up and running:

March 4, 2019

Hello There, Mini Cheetah

Now that this is officially out I can finally put up something about what I've been doing for the past 2 years.

Photo Credit: Bayley Wang

This project has been a continuation of the Hobbyking Cheetah project, but for research rather than out of my pocket.  Here's my thesis on the actuator and robot design.  I designed and built all the all the hardware, both mechanical and electronics (with some design and fabrication help from Alex), and Jared wrote pretty much all the software and high-level control running the robot, including the Convex Model Predictive Control for locomotion.

The design principles behind the robot are very similar to Cheetah 3:  High torque electric density motors, low gear ratio, efficient transmissions to minimize friction and reflected inertia, lightweight legs with motors placed to minimize leg inertia.  A few things are different, though.  Mini Cheetah has 12 identical modular actuators with built-in motor control, gearbox, and support structure.   The electric motors are off-the-shelf and very cheap, unlike Cheetah 3's custom designed motors.  Mini Cheetah has even more range of motion than Cheetah 3 at the hip, so it's able to point it's legs completely sideways.  This means that it should be possible to make the robot land feet-first regardless of what orientation it falls in.  

I finished the hardware for the robot last May, but the publication cycle is kind of slow, so I wasn't  able to put up any info about it before.  There's going to be a paper about the robot in ICRA 2019 in Montreal, so the robot and I will both be there.


Jared and I (mostly Jared) did the backflips for our final project for Underactuated Robotics last spring.  There are some more details about how it works in the thesis, but basically we did a nonlinear trajectory optimization offline, and just played back the resulting joint torques on the motors, with a little bit of joint position control to the optimized trajectory on top.  Our simulation of the dynamics was accurate enough that it worked on the first try in the hardware.

Here are the videos from back in May.  At the time we weren't set up for backflips completely untethered and wireless.  Since then we've adjusted the flip trajectory a bit, made the landing less bouncy and more reliable, and gone fully wireless.

February 9, 2019

Big Dyno Up And Running

A long overdue progress update on the big dyno.

The torque sensor got new electronics Bayley designed, which have a built in power supply, A/D, and micro, to convert the signal to digital as quickly as possible.  I made a nice box to house all the electronics:

The torque sensor mounted, with the absorber on the right and the motor under test on the left:

To power and control the test motor, there's a bank of terminals for DC input and motor phase output, if you want to drive the test motor with the Prius inverter built into the dyno.  And quick disconnect fittings for water cooling the test motor:

Speaking of inverters, here's a look under the sheet metal covers:  Underneath are two Prius "bricks" mounted to a big waterblock.  Although the Prius power modules have two 3-phase inverters each, we had to use two of them so we could separate out the DC link current measurement for the inverter powering the test motor.

The dyno has an internal Mac Mini with linux on it, runing a fork of my dyno software.  Thanks to the cross-platform-ness of Python and Qt, all I had to do was install all the libraries and change the "COM" serial ports to "/dev/ttyUSB" and it just worked.  Yes, that's the "Hybrid Synergy Drive" plaque from the Prius power module covering up the Apple logo:

To go with the computer, it has a built-in monitor which can be raised and lowered for storage:

The absorber and all the sensors work, and we've helped the FSAE team do a little bit of testing on their new motors.  Next up is populating the second half of the control electronics so the dyno can control the test motor.  Then we'll be able to do a full optimization of the Sonata HSG and get even more power out of the big kart.

December 21, 2018

Planar Magnetic Headphones, Part 3: Finishing Up

Over a year later, I've finally finished up the planar magnetic headphones.  They even sound decent.

Here's the rest of the build log:

The change that got them from sounding terrible to sounding tolerable was reducing the tension in the membrane as much as possible.  The first few test versions were intentionally tensioned, which turned out to be a bad idea, causing weird notches and resonances in the frequency response.  The new membrane is glued to a frame under no tension.  I also switched to an even thinner flex pcb material, Pyralux AC 091200EV, which has a 12 micron kapton layer, 9 micron copper layer, and no adhesive.

After the toner transfer.  Getting the toner transfer right without wrinkling the material was tricky:

After etching.  The driver came out to about 50 ohm resistance.

This time I glued the membrane to a fiberglass frame.  To get the PCB flat without tension, I stuck it down to a metal plate by wetting the plate.  I forgot to take a picture of these assemblies, but here's one of the original PCBWay flex pcbs glue to an identical frame:

For the final headphones I'm using a single sided magnet array - I first built a single driver, and did experimenting with both single and double arrays, and qualitatively could not hear any significant difference between the two.  And a single array is about half the effort of the double array.  Here's the test-driver, with a plastic back, and the first ultra-thin board I was able to successfully etch:

I machined two new aluminum backs, and stole the headband from a cheap pair of headphones.  Maybe eventually I'll get around to making a nice headband, but for the time being I'm not particularly interested in building headbands:

On a dummy head:

They sound decent.  They're a bit bass-heavy, and require some equalizing, but once that's done I quite like them.

December 1, 2018

CORE Outdoor Power PCB Motor Teardown

A company called Core makes some unusual electric lawn tools based on axial flux PCB motors.  I first heard about them several years ago, and finally picked up one of the motors to take apart.

Here's the weed-wacker.  I just got the head of the tool- they make a base called the "drive unit", which holds the battery and electronics.  Each head has its own motor:

The motor's right at the end, direct driving the spool of wire:

Removed from the stick.  Three board-mount faston terminals for phase power, and another connector for hall sensors.

The motor was a huge pain to take apart, as it is pretty much entirely held together with retaining compound.  With a hammer and some screwdrivers I was able to pull off the back cover:

The shaft and front bearing were pushed out of the other half of the housing with a gear puller.  There's a ring of thermal pad on one side of the board:

It's a dual-rotor design, with the PCB stator sandwiched between the two halves of the rotor.  The rotor is held together by even more retaining compound:

To pull off half of the rotor, I gripped it in a small lathe chuck and used  a gear puller to pull it off the shaft:

First a closer look at the stator.  The board had at least 4 layers, and probably has 4 oz copper - it's very dense feeling. 

Some observations and thoughts:

- Each phase has 4 4-turn coils, and the 3 phases are overlapped.   This is a full-pitch winding.
- This winding pattern results in a large area of end-turn, so the actual "active" area of the stator is unfortunately small - less than half the area of the PCB actually produces torque.
- The traces neck down and become thinner within the active area.  Presumably this is to reduce eddy current losses from the magnets swinging over the traces. 
- The hall sensors are through-hole halls, surface mounted sideways into cutouts in the PCB, so they add almost no additional thickness.
- The empty space around the edges of the board are filled with radial strips of copper.  I'm not really sure why.  Maybe for thermal reasons?  The edge of the board is heatsinked to the aluminum housing through a ring of thermal pad.

Here's the patent for reference.

The two rotors have single-piece magnets magnetized with 4 pole-pairs:

One of the reasons it's all held together with retaining compound is that the bearing bores in the castings aren't even post-machined.  The bearing is just glued straight into the rough, tapered hole in the casting:

Back-EMF is pretty sinusoidal.  An FFT reveals a little bit of 5th harmonic.  Flux linkage is ~.0044, for a torque constant of .0264 N-m/A (peak phase amps).  Line-to-line resistance is 125 mOhms. This gives a motor constant of .086 N-m/sqrt(watt).

Over all it's not a particularly high performance motor for it's size or weight.  There's a lot of dead space, and a very heavy, high-inertia rotor. 

September 29, 2018

Controlling Phase Current Harmonics with FOC and AFC

I was playing with my thermal test motor, and noticed that at high speeds, some pretty significant harmonics showed up in the phase currents.  Here's an example with an obvious 5th harmonic - I'm commanding 20A, peak, and the harmonics add a couple amps of ripple on top of that:

This ripple shows up because there's a 5th harmonic in the motor's flux linkage (and therefore back-emf).  From the perspective of the field-oriented control, these harmonics at the phases show up as AC disturbances on the D/Q voltages.  The two current loops will try to squash these disturbances, but it's ability to do so depends on the current loop bandwidth and the magnitude of the disturbances.  Since both the magnitude and frequency go up with motor speed, even  really high-bandwidth current loops have trouble keeping up.  For example, in the scope shot at the top, the current loop had around 2 kHz closed-loop bandwidth.  The motor's electrical frequency was 300 Hz, so the harmonic (in D/Q) was at 1.8 kHz, so the controller couldn't quite reject it.

So how exactly do harmonics in the flux linkage at each phase show up when transformed into D and Q?  In the sinusoidal flux linkage case with no position-varying inductance, the motor's dynamics in the D/Q frame is exactly the classic voltage equations:

$$V_{d} = Ri_{d} + L\frac{di_{d}}{dt}  - \omega Li_{q}$$
$$V_{q} = Ri_{q} + L\frac{di_{q}}{dt}  + \omega Li_{d} + \omega\lambda$$

Where \(L\) is the synchronous inductance, \(R\) is the synchronous resistance, \(\omega\) is the (electrical) angular velocity, and \(\lambda\) is the peak flux linkage of one phase.

In the equations above, the flux linkage of a single phase \(\lambda_{p}\) is assumed to be of the form \(\lambda_{p} = \lambda \cos(\theta)\), where \(\theta\) is the rotor angle.  Other phases will have an offset of \(\pm \frac{2\pi}{3}\).

If the flux linkage has a fifth-harmonic in it, for a single phase it looks like \(\lambda \cos(\theta) + \lambda_{2}\cos(5 \theta) \).  With the fifth harmonic in flux linkage, the voltage equations in D and Q become:

$$V_{d} = Ri_{d} + L\frac{di_{d}}{dt}  - \omega Li_{q} + 5\omega\lambda_{2}\sin(6\theta)$$
$$V_{q} = Ri_{q} + L\frac{di_{q}}{dt}  + \omega Li_{d} + \omega\lambda - 5\omega\lambda_{2}\cos(6\theta)$$

(I have no interest in doing tedious trig and algebra by hand, so I wrote a Mathematica script to do the transforms analytically for me.)

That's interesting  - a fifth harmonic at the phases shows up as a 6th harmonic through the transforms.  It turns out, a 7th harmonic at the phases also shows up as a 6th harmonic through the transforms - just with both positive sine and cosine components, rather than positive sine negative cosine.  11 and 13 show up as 12, and so on.  Also (although this may be more intuitively obvious), odd multiples of 3 disappear through the transforms.

Just to confirm, here's the output when plugging a fifth harmonic into my motor simulator, using the the other motor parameters from the U12 and same controller gains as the hardware.

Phase currents:

D and Q axis currents, as well as rotor angle.  You can count the 6 periods in the AC component of the current per electrical cycle.

So what if you want to control out those harmonics?  Or add in harmonics of your own, for some reason?  If you know how the voltages of the harmonics appear on the D and Q axes, theoretically you could feed-forward them out, but there's a more interesting way to closed-loop deal with them, without knowing the specific amplitude of phasing of the harmonic, just it's frequency.

Introducing Adaptive Feedforward Cancellation.  This is a weird technique which was briefly introduced at the very end of 2.14, which has the seemingly magical property of giving your controller infinite gain at a specific frequency - i.e. it will perfectly track a signal at that frequency, or perfectly reject a disturbance at that frequency.  AFC is probably most well-known for its use in things like hard-disk read heads, which have to follow very fine tracks of data on the hard drive platters.  Since the platters have some runout as they spin, the read-head has to track in the presence of periodic disturbances.  Interestingly, AFC is a lot like field oriented control: FOC gives your controller infinite gain at the fundamental electrical frequency of your motor, letting you track sinusoidal currents of much higher frequency than the bandwidth of your current controllers.

Here's a block diagram of how AFC works, taken from this paper.  \(r(t)\) is the reference (current, in my case), and \(d(t)\) is a disturbance (back-emf harmonics, in my case), and \(y(t)\) is the control effort (D or Q voltage in my case).

And here's how the AFC(s) are integrated into the control loop.  The red box is your normal feedback controller (in my case the PI current controllers for D and Q current).

The AFC works kind of like a pair of integrators on the sine and cosine components of the periodic part of the error.  If there's zero error \(e(t)\), the AFC outputs the sum of \(\cos(\omega t)\cdot (\text{cos integral})\) and \(\sin(\omega t)\cdot (\text{sin integral})\), which is some sinusoidal thing.  If there is some error, the portion of it at the same frequency as \(\omega\) doesn't average to zero when multiplied by the \(\cos(\omega t)\) or \(\sin(\omega t)\)'s before the integrators.  The integrators then integrate up to whatever value drives the  \(\omega\) frequency component of the error to zero.

The slight modification to get this to work for motor currents is changing the \(\omega t\) to be \(\theta\), the electrical angle of the motor.  The periodic disturbances are not a fixed frequency in time, they are a fixed frequency in state. 

This basically just works.  Implementing a 6th harmonic AFC on both the D and Q current loops completely eliminates the 5th harmonic in phase currents.  Here's the simulation result:

Phase currents:

D/Q currents:

And here's what the integrators in the AFC do:

There's one simple addition to the controller which makes it behave a little better.  The magnitude of the harmonic voltage is proportional to how fast the motor is spinning - so if the motor is accelerating, the magnitude is increasing, and the AFC integrators are constantly playing catch-up.  This means that at a constant acceleration, there will be a constant error in the AFC:

The AFC still does a pretty good job even at high acceleration, but how well it does depends on how large the AFC gain is.  Below you can see the integrators ramping up with speed:

The simple fix for this is to divide the error by angular velocity before integrating, and then multiply by angular velocity at the end of the AFC calculations.  This way, the AFC integrators integrate up to the harmonic in the flux linkage, which is constant, rather than the back-emf, which is proportional to speed.  Honestly, this change doesn't make a huge difference, because the time constant of the integrator can be quite fast compared to how fast the motor can accelerate, but it's kind of cool none the less:

Now the integrators just converge to constants, even as the motor accelerates:

Pseudo-code implementation:

h = 6.0f;  // Harmonic number
k_afc = 100;  // AFC Gain - larger gain converges faster

s = sin(h*theta);
c = cos(h*theta);

i_error = i_ref - i;
if(abs(thetadot) > 1) //avoid dividing by zero
 afc_cos_int += dt*i_error*k_afc*c/thetadot; 
 afc_sin_int += dt*i_error*k_afc*s/thetadot;
afc_out = thetadot*(s*afc_sin_int + c*afc_cos_int);

i_error = i_ref - i + afc_out;  //Now use this error in the PI current loop.

And here it is after implementing on the hardware:

The AFC behaves a little strangely when the voltage saturates (causing higher harmonics to show up), and I haven't put any effort into fixing that, but for most operating points it works quite well.

Here's an FFT of the phase current at the beginning of the post.  There are significant components at 5, 7 and 11 times the electrical frequency:

And here's the FFT after adding an AFC at 6 times the electrical frequency:

In addition to canceling out harmonics, you can use this technique to add in harmonics of your own, or track periodic current references.

For example, a typical way to cancel out cogging torque is to add in an AC component to your Q-axis current reference which tries to cancel out the cogging torque.  The controller's ability to track this reference depends on how fast the motor is spinning, and its bandwidth.  But if you add in an AFC at the cogging frequencies, the AFC just takes care of tracking the cogging current.

Or if you want to add certain harmonics at the phases, you can do so by taking the dq0 transform of the harmonics you want, then adding those to your D/Q current references with an AFC to track them.  Say, for example, you wanted your phase currents to be more trapezoidal, you could take the transform of the trapezoidal waveform, use that as your D/Q current references, and add AFCs at the first few Fourier coefficients in order to track the reference at high speed.

August 21, 2018

Thermal Testing, Round 2

I made a cover for the small motor controller, which serves as both a top-side heatsink for the MOSFETS and gate driver, and (more importantly) as a clamping mechanism to keep the PCB flat and evenly pressed against the heatspreader beneath it.

In the picture below, the heatspreader is on the left, and cover on the right:  The bar across the bottom sits on top of the FETS, the square in the middle sits above the DRV chip.  I left a 0.1mm gap between the FET-bar and the top of the FET package, which gets filled by some .5mm thermal pad.  This allows for some variation in the height of the transistors once they're soldered down, and the compression of the thermal pad means there's a fair amount of pressure pushing down on the top of each transistor.

Motor Controller Sandwich:

The two little holes let through light from the two indicator LEDs on the board.

I made a Special Motor Controller Cover of Science to help out with thermal camera-ing.  It has a 1.5mm hole drilled above the center of each transistor, so that I can see through the cover with the camera.  I stuck it on the same controller and motor I used for the last round of testing.  I was able to use a needle to cut away the circle of thermal pad at the bottom of each hole, so you can see straight through to the tops of the FET packages.

Here's an example image You can see the transistors clearly through the holes.  The bright spot on the cover where point Sp2 is is where I applied some matte black paint to get the emissivity of the aluminum cover closer to 1.

The cover was extremely effective.  Here's a 30A test.  By the time the motor windings reached 70 C, the FETs were all under 48 C, and there was only a 1.6 C difference between the hottest and coldest transistors.  For reference, in my last tests without the cover, by the time the motor reached 70 C at 30 amps, the FETs were between 67 and 72 C.

Going a little harder, here's after running 33 A until the motor windings reached a rather uncomfortable 80 C.  There's a 5 degree drop between the top of the transistor and the surface of the cover.

That's all for now.