Inside Haddington Dynamics and Their Groundbreaking Dexter Robot Arm(blog.hackster.io)
blog.hackster.io
Inside Haddington Dynamics and Their Groundbreaking Dexter Robot Arm
https://blog.hackster.io/inside-haddington-dynamics-and-their-groundbreaking-dexter-robot-arm-db9a7a965445
7 comments
The encoder is my very favorite part of the robot. I've written a page about how it works in detail:
https://github.com/HaddingtonDynamics/Dexter/wiki/Encoders
In summary, it's a hybrid of digital and analog. To the standard quadrature encoder slit counter, we add sensing of the light levels as each slit covers or obscures the phototector. So we get e.g. 200 slits times 8096 light levels inside each slit equals over a million positions. It's so much data that we need an FPGA to process it fast enough. You can breath on the arm and it can sense it.
There is a constant battle between writing for sales/investment and writing for technical disclosure. Sometimes we get some chocolate in the peanut butter. It's certainly not phony, and I'm happy to answer any questions that anyone has about the real technical abilities.
The $2,999 price was for the Dexter 1 which was available at the time the article was written. The current generation is the Dexter HD which is more capable, much easier to build, and (as you noticed) more expensive.
In summary, it's a hybrid of digital and analog. To the standard quadrature encoder slit counter, we add sensing of the light levels as each slit covers or obscures the phototector. So we get e.g. 200 slits times 8096 light levels inside each slit equals over a million positions. It's so much data that we need an FPGA to process it fast enough. You can breath on the arm and it can sense it.
There is a constant battle between writing for sales/investment and writing for technical disclosure. Sometimes we get some chocolate in the peanut butter. It's certainly not phony, and I'm happy to answer any questions that anyone has about the real technical abilities.
The $2,999 price was for the Dexter 1 which was available at the time the article was written. The current generation is the Dexter HD which is more capable, much easier to build, and (as you noticed) more expensive.
Seems like a device of this capability will be an easy sell to anyone who is going to use it to make money. It only seems expensive to a hobbyist.
So, actually, servos with >1 million counts per rev are not that uncommon. They are now standard.
Example: the mr-j4's on my CNC machine have 22 bit encoders (4 million counts per rev) and are 4+ years old at this point.
These are fairly standard equipment, and not that expensive as servos go (300 a motor, 300 a servo drive, in bulk).
It will happily position repeatedly at to any of those counts, and it's repeatable to better than 5 microns (i am only using equipment good to about 5 microns because that's what i need. I have no urge to go set up the laser interferometer to measure this better, but i have no doubt it is repeatable to better)
Yasakawa's sigma7 series are 24 bit encoders (so 16 million count per rev).
Basically current generation servos are easily in the millions per rev count.
(and since the robot are in turn made from these servos, ...) To be fair to haddington: this is a pretty wild leap from the last generation of servos in most cases.
For comparison, the melservo-j3 was only an 18 bit encoder (262,114 count per rev), and the yasakawa sigma 5's were 20 bit (1 million count per rev)
So each got better by a factor of 16 in the newest generation
Example: the mr-j4's on my CNC machine have 22 bit encoders (4 million counts per rev) and are 4+ years old at this point.
These are fairly standard equipment, and not that expensive as servos go (300 a motor, 300 a servo drive, in bulk).
It will happily position repeatedly at to any of those counts, and it's repeatable to better than 5 microns (i am only using equipment good to about 5 microns because that's what i need. I have no urge to go set up the laser interferometer to measure this better, but i have no doubt it is repeatable to better)
Yasakawa's sigma7 series are 24 bit encoders (so 16 million count per rev).
Basically current generation servos are easily in the millions per rev count.
(and since the robot are in turn made from these servos, ...) To be fair to haddington: this is a pretty wild leap from the last generation of servos in most cases.
For comparison, the melservo-j3 was only an 18 bit encoder (262,114 count per rev), and the yasakawa sigma 5's were 20 bit (1 million count per rev)
So each got better by a factor of 16 in the newest generation
and FWIW:
The near state of the art that i'm aware of is the renishaw resolute:
https://www.renishaw.com/en/resolute-absolute-encoder-system...
Being rotary encoders, they are measuring accuracy/etc in arcseconds. For reference, 1 arcsecond is ~0.0003 degrees (1/3600th of a degree). So the +-5 arcsecond on the largest rings is 0.0015 degrees, and the +-0.5 arcsecond on the smaller is .000015 degrees.
Both are fairly ridiculously small angles.
To give a sense of scale of this error. On a circle with a radius the size of the earth, 1 arcsecond would only make an arc 30 meters long.
On a circle with a 1 meter radius, it's about 5 microns of arc length per arcsecond.
So yeah, you can do pretty well now in encoders :)
The near state of the art that i'm aware of is the renishaw resolute:
https://www.renishaw.com/en/resolute-absolute-encoder-system...
Being rotary encoders, they are measuring accuracy/etc in arcseconds. For reference, 1 arcsecond is ~0.0003 degrees (1/3600th of a degree). So the +-5 arcsecond on the largest rings is 0.0015 degrees, and the +-0.5 arcsecond on the smaller is .000015 degrees.
Both are fairly ridiculously small angles.
To give a sense of scale of this error. On a circle with a radius the size of the earth, 1 arcsecond would only make an arc 30 meters long.
On a circle with a 1 meter radius, it's about 5 microns of arc length per arcsecond.
So yeah, you can do pretty well now in encoders :)
> Where did they get a shaft encoder with over a million counts per rev?
they use about 200 tick encoders, but instead of measuring on or off they measure light intensity. Because each encoder slot has a bit of variation they get absolute positioning.[0] This is what their FPGAs are for.
As for the reason this robot is so expensive, the inconel harmonic drive gear boxes are probably a good portion of the cost. Harmonic drives aren't made in big production runs. The assembly instructions are somewhat horrifying and involve quite a bit of manual labor.[1] There's a lot of cutting carbon fiber rod and gluing it into the 3d printed bits. The horrifying bit is that you have to glue stuff to the motors and it appears that some assemblies aren't exactly made to come apart.
[0]https://github.com/HaddingtonDynamics/Dexter/wiki/Encoders [1]http://hdrobotic.com/dexter-v1-assembly-manual/
As for the reason this robot is so expensive, the inconel harmonic drive gear boxes are probably a good portion of the cost. Harmonic drives aren't made in big production runs. The assembly instructions are somewhat horrifying and involve quite a bit of manual labor.[1] There's a lot of cutting carbon fiber rod and gluing it into the 3d printed bits. The horrifying bit is that you have to glue stuff to the motors and it appears that some assemblies aren't exactly made to come apart.
[0]https://github.com/HaddingtonDynamics/Dexter/wiki/Encoders [1]http://hdrobotic.com/dexter-v1-assembly-manual/
> Where did they get a shaft encoder with over a million counts per rev?
They didn't. They use a wheel with fewer counts per rev, read an analog value, and interpolate. It's like micro-stepping a stepper motor.
https://hackaday.com/2018/11/15/an-in-depth-look-at-dexter-t...
They didn't. They use a wheel with fewer counts per rev, read an analog value, and interpolate. It's like micro-stepping a stepper motor.
https://hackaday.com/2018/11/15/an-in-depth-look-at-dexter-t...
Oh, like a daisy-wheel printer.[1] The daisy wheel is mostly driven as a stepper motor, but fine positioning uses an analog sensor.
[1] http://laughtonelectronics.com/oldsite/comm_mfg/commercial_d...
[1] http://laughtonelectronics.com/oldsite/comm_mfg/commercial_d...
The difference between hobby/research robot and similar industrial robot is the cost of maintenance and uptime.
If you run cheap robot continuously, it wears down quickly.
If you run cheap robot continuously, it wears down quickly.
The FPGA enables higher sampling rates, which allows for the finer resolution when paired with a lookup table which contains systematic error.
RLS make magnetic encoders going up to 600k CPR (ring with dozens of magnet poles + multiple sensing elements): https://www.rls.si/en/products/rotary-magnetic-encoders
That would still need very accurate homing switches though.
That would still need very accurate homing switches though.
Yasakawa has some very neat batteryless absolute encoders (so never need rehoming) in the latest sigma 7.
http://www.e-mechatronics.com/download/datas/catalog/cheps80...
They use the power generated by the motor spinning to store the position.
http://www.e-mechatronics.com/download/datas/catalog/cheps80...
They use the power generated by the motor spinning to store the position.
> Where did they get a shaft encoder with over a million counts per rev?
The article says they put the encoder directly on the motor, before gearing it down. So the effective resolution after gearing is magnified. But the encoder moves much faster accordingly, hence the need for higher speed / fpga decoding.
The article says they put the encoder directly on the motor, before gearing it down. So the effective resolution after gearing is magnified. But the encoder moves much faster accordingly, hence the need for higher speed / fpga decoding.
No, the article says "Dexter’s optical encoders are placed directly on each joint, and have resolutions that aren’t possible with any other control system." The usual approach is an encoder on the motor, before the gear train. They probably have one too, so they can see what the motor is doing vs. the joint. Remember, none of this is rigid under load.
There's just one encoder per joint. The encoder disk is 3D printed and the cruft from the print is used to refine the encoder ticks. Instead of fighting noise they save a map of all the systematic noise and then can use a lookup table to localize.
*I went to Vegas and built an arm with the founders.
*I went to Vegas and built an arm with the founders.
Does this introduce risk of future noise (fluff, chad, whatever: it's industrial applications == dirt == hostile environment) being misread, and require recalibration?
Not saying wontworkbecause. I think this calibration trick is very good and I am imagining it's applications to other things eg more precise disk head placement from noise maps of the interblock separation reads
Not saying wontworkbecause. I think this calibration trick is very good and I am imagining it's applications to other things eg more precise disk head placement from noise maps of the interblock separation reads
Ya recalibration is definitely a necessary aspect of this approach if the optical reading changes. For example if IR light levels change.
Which is guaranteed with aging.
Encoders at the joint only. Motors are steppers, so we have some idea where they are but that is open loop.
I've been following these guys for a while and I really like their approach. In our lab we've done a lot of work to improve control of cheaper BLDC motors (low gearing) to mimic human control and compliance, and these guys are using similar methods to achieve accuracy. Most engineers just decrease tolerances to improve accuracy but using a blunter instrument and making it accurate with good software is an approach that leads to much much cheaper systems and robots. Drones are a good example, quadcopters are inelegant and somewhat inefficient flying machines but their cheap and blunt hardware comes alive with good control theory and software.
> There are no billionaire tech entrepreneurs riding Segways to meetings with angel investors, and there are no corporate PR-approved mission statements on the wall. Instead, you’ll find a small team of brilliant engineers who are using passion and innovative technology to design a very special robot called Dexter.
That's, uh, that's exactly what a state-of-the-art robotics lab is supposed to look like. I used to work in one; place was a mess. Then they closed some big sales, moved to a shiny new facility -- and promptly got worse, that is, worse if your goal is to actually work on cool robots rather than edit elaborate specs on theoretical robots to get Major Defense Client to sign the contract.
That's, uh, that's exactly what a state-of-the-art robotics lab is supposed to look like. I used to work in one; place was a mess. Then they closed some big sales, moved to a shiny new facility -- and promptly got worse, that is, worse if your goal is to actually work on cool robots rather than edit elaborate specs on theoretical robots to get Major Defense Client to sign the contract.
Haddington's arm is 5-axis, which is not enough to put a tool in an arbitrary position and orientation. To put a tool in an arbitrary position (3 DOF) and orientation (3 DOF), a 6-axis arm is needed. If their encoder technology is so good, why aren't they making a 6-axis arm and going right up against Universal Robotics?
I'm not certain about this, but I believe that their arm does not have a final rotary actuator on the end of the arm. If the arm is used to hold a drill, for example, that does not matter.
I don't know about the robotics community on hn, but I use the Universal Robotics robot arms for sensor testing. Sensor Testing requires 6 DOF. Current Universal Robotics arms go for ~50k, and for that you get a highly repeatable robot arm, which is also safe for operation near humans, and with a good UI.
What is so hard about the final joint, Haddington? If this is a product decision, to go with a 5 axis first to get to market first, I disagree. At this moment, 5-axis robots seem more useful to manufacturing than to roboticists and robotics hobbyists. However, I would expect manufacturing to have long acquisition cycles, and to be more willing to select an expensive, yet proven, robotics company, than to take a chance on a cheap upstart, whatever their technology.
However, if you can get a 6DOF arm to within the budget of a university robotics class or individual hobbyist, who will be willing to take a chance on an unproven robotics startup, then these people will build word of mouth about your design. Teaching the math of robotic manipulation requires a 6+ DOF robotic arm. Students who learned on a Haddington arm will go into industry knowing the Haddington interface. Furthermore, it's much more likely for a $3k arm to be within the discretionary budget of an engineering team than a $50k one.
6DOF or bust, I just don't get it.
I'm not certain about this, but I believe that their arm does not have a final rotary actuator on the end of the arm. If the arm is used to hold a drill, for example, that does not matter.
I don't know about the robotics community on hn, but I use the Universal Robotics robot arms for sensor testing. Sensor Testing requires 6 DOF. Current Universal Robotics arms go for ~50k, and for that you get a highly repeatable robot arm, which is also safe for operation near humans, and with a good UI.
What is so hard about the final joint, Haddington? If this is a product decision, to go with a 5 axis first to get to market first, I disagree. At this moment, 5-axis robots seem more useful to manufacturing than to roboticists and robotics hobbyists. However, I would expect manufacturing to have long acquisition cycles, and to be more willing to select an expensive, yet proven, robotics company, than to take a chance on a cheap upstart, whatever their technology.
However, if you can get a 6DOF arm to within the budget of a university robotics class or individual hobbyist, who will be willing to take a chance on an unproven robotics startup, then these people will build word of mouth about your design. Teaching the math of robotic manipulation requires a 6+ DOF robotic arm. Students who learned on a Haddington arm will go into industry knowing the Haddington interface. Furthermore, it's much more likely for a $3k arm to be within the discretionary budget of an engineering team than a $50k one.
6DOF or bust, I just don't get it.
Haddington's arm is 5-axis, which is not enough to put a tool in an arbitrary position and orientation. To put a tool in an arbitrary position (3 DOF) and orientation (3 DOF), a 6-axis arm is needed.
Their gripper option adds another axis.
Their gripper option adds another axis.
It sounds impressive, but "FPGA supercomputer"? It's a few optical shaft quadrature decoding state machines and feedback to motor controllers. Describing that as a 'supercomputer' suggests there's a lot of marketing bs flying around.
Still if its resolution really is 50 micron it should be able to, eg, mill useful pcbs with the right head.
Still if its resolution really is 50 micron it should be able to, eg, mill useful pcbs with the right head.
The encoder / motor controller logic is a bit more complex than that. Because the encoders are analog / digital hybrid they generate far more data than standard quadrature encoders. The FPGA is doing huge table lookups, ATAN2, PID, Dithering the motor driver, and all of that at 5 MHz end to end, so we have a 200ns response.
How huge is huge? The rest of the chain sounds pretty straightforward, even at 5MHz. That's not to say that an FPGA is wrong for the application, but nothing in that list screams "must use FPGA" to me.
However, the whole program screams "get to market, then iterate" to me, which is a very effective way to move forward with a product like that.
Note: I've shipped products doing continuous signal processing on DSP, GPU, CPU, SoC, and FPGA. (MCU eventually... Give me time.) I did the mental swag on the described processing chain...
Out of curiosity, how big is your FPGA?
However, the whole program screams "get to market, then iterate" to me, which is a very effective way to move forward with a product like that.
Note: I've shipped products doing continuous signal processing on DSP, GPU, CPU, SoC, and FPGA. (MCU eventually... Give me time.) I did the mental swag on the described processing chain...
Out of curiosity, how big is your FPGA?
It could probably be handled by the realtime cores on the BeagleBone Black's SoC, or perhaps by an XMOS part, but if someone already knows how to develop for FPGAs there's not much of an incentive to try to shoehorn the task onto a microcontroller.
Anything that involves reading from moderately-to-insanely fast ADCs is a natural FPGA application, especially if multiple channels are involved.
Anything that involves reading from moderately-to-insanely fast ADCs is a natural FPGA application, especially if multiple channels are involved.
I hear that. I'm working on multi-channel ADC hardware right now, and there is no way that we're going to use an FPGA (per unit cost).
At my last gig, we used FPGA's for several low-volume products.
At my last gig, we used FPGA's for several low-volume products.
One of the founders wrote a FPGA language. I suspect it's the tool the team knows the best, but may not be the most price optimized. I think that's fine for now, especially since it's still one of the most affordable arms available.
> A supercomputer is a computer with a high level of performance compared to a general-purpose computer.
https://en.wikipedia.org/wiki/Supercomputer
By this definition any FPGA that has been correctly programmed is a super computer relative to a PC or non-parallel circuit performing the same task.
Couldn't you mill a PCB a lot cheaper with a 3 axis machine?
Edit: On the other hand, it would be nice to be able to probe a complex surface.
https://en.wikipedia.org/wiki/Supercomputer
By this definition any FPGA that has been correctly programmed is a super computer relative to a PC or non-parallel circuit performing the same task.
Couldn't you mill a PCB a lot cheaper with a 3 axis machine?
Edit: On the other hand, it would be nice to be able to probe a complex surface.
Your definition is way too broad because it depends on the specific computer you're using as a comparison. Almost everything is a supercomputer compared to e.g. a raspberry pi zero.
I believe they have a version called electron which is doing smd placement.
The reason encoders go on the motor and not at the far end of the gears is because inexpensive gearing has backlash, which will cause the motor to dance back and forth to keep the encoder at a specific point.
So how are they solving the gearing problem? Harmonic drives are expensive and patented.
So how are they solving the gearing problem? Harmonic drives are expensive and patented.
1. The encoders on the Dexter arm are at the joints because that gives us the ability to compensate for backlash in the drive system.
2. The motor dance issue happens because the feedback loop on most control systems isn't fast enough. With the FPGA, our control loop is 5 MHz / 200nS so we can correct before an oscillation has time to happen.
3. Our encoders give more than 1 million CPR (clicks per revolution) so we can sense an error before it has time to become an issue that might lead to oscillation.
(2 and 3) are the real magic of the arms design
4. Yes, Harmonic drives are hell. And they are not actually free of backlash, it's just "soft" e.g. they bend slightly. At this point, they are the best solutions, but we have prototypes of a cycloidal drive that costs less (in quantity) and gives us better performance. There are issues to work out, especially related to smoothness and wear, but we are making progress on it.
2. The motor dance issue happens because the feedback loop on most control systems isn't fast enough. With the FPGA, our control loop is 5 MHz / 200nS so we can correct before an oscillation has time to happen.
3. Our encoders give more than 1 million CPR (clicks per revolution) so we can sense an error before it has time to become an issue that might lead to oscillation.
(2 and 3) are the real magic of the arms design
4. Yes, Harmonic drives are hell. And they are not actually free of backlash, it's just "soft" e.g. they bend slightly. At this point, they are the best solutions, but we have prototypes of a cycloidal drive that costs less (in quantity) and gives us better performance. There are issues to work out, especially related to smoothness and wear, but we are making progress on it.
I have a Dexter. They have harmonic drives from China which were deemed out of spec for normal usecases but are fine for the 3kg realm. Their latest version is running a custom cycloidal gear system.
July 2014. The article is a little out of date to have a title of Groundbreaking.
The article misses the point. The purpose of the high resolution sensing is not positional control. It's for force feedback. Here's the spec page for the robot, with more info.[2] (But too much hype. Seeing "In the loop, On the loop, Out of loop, Big Data Roadmap" on a data sheet makes me think "phony".)
Also, it's not $2,999. It's $4,888 for the base kit, and $9,800 complete and assembled. You don't have a usable robotics system at that point - you need another computer to control it. The FPGA is just motor control. Still, that's better pricing than Universal Robotics, which starts around $35K for a mechanically comparable system.
There's no reason these things should cost more than many cars. They're not that complicated. It's mostly volumes being low. That problem took down Rethink Robotics. I have a $400 3-axis uArm on my desk. This is a low end robot with RC servo grade motors. I built a six-axis force feedback end effector out of a Space Navigator for it. Controlled by ROS. The functionality is there; the precision and useful life are not.
[1] https://www.mouser.com/new/broadcom/avagoAEAT9000/ [2] http://hdrobotic.com/tech-hub/