It turned out just enough of the ancient DT700's were acquired over the years to make exactly 8 legs. 1 would have to be rewound & the exact number of turns is unknown. It seemed to be 19 turns in a delta. The ESC department has 5 Hobbyking 35A's, 1 Hobbyking 18A, 1 Castle creation 35A, 1 Align 35A. The last 2 are based on the 8051, so quite hard to reprogram.
They're amazingly still sold. They performed much better than the smaller motors. They could jump the load with only 12V. With active cooling, encoder feedback, reducing the lever size, they could probably just manage a slow bounding movement.
Active cooling has been deemed essential for any application. Even with active cooling, overheating still makes it impossible for them to hold the torque longer than a second. The mane problem with continuing is finding encoders for feedback.
Call it ignorance about magnetic theory, but for jumping movement, 9 turns ended up better than 50 turns. Where the 50 turn couldn't jump at all, the 9 turn jumped the 140g load quite efficiently with very little heating, at 20V. Surprised the maximum starting torque as well as the maximum speed increased, since one would think the 2 would trade places. Holding movement suffered, since it could no longer hold the load without instantly overheating.
The power supply showed 8A being used for jumping & 3A used for holding. The holding current couldn't be fine tuned below that. Software PWM isn't precise enough. There's a way to theoretically determine the optimum turns count, but prototyping has proven faster.
Decided to use active cooling to reduce the risk of a burnout. A scaled down version of the Minotaur's bounding movement is certainly possible. It would start as a single leg with 2 wheels, to predict performance. It wouldn't stand up when not jumping.
There was 1 rewound motor left from 2014. Even though the requirements were never met, it was worth documenting what it did. Drove it at 16.8V to simulate a 4S battery. Set the load to 140g. Hard coded a jumping movement which powered it for 250ms at full power & a holding movement which powered it at 50% power for 1700ms. This managed to keep it from burning out, but it heated up quickly. Care had to be taken to let it cool down & it could never hold the load continuously.
It never got the load to a parabolic trajectory, but it was an open loop programmed as close as possible to the ideal acceleration. A closed loop with encoder feedback might have gotten more acceleration to it. The bare motor was 70g, so the load simulated the motor lifting itself + any required accessories.
Rewinding the motor may have made it more optimal for holding loads at the expense of jumping movements. It might be better off with a higher KV.
It was around this time the lion kingdom finally read more of kaloucheThesis.pdf & discovered the Ghost Robotics Minotaur. That thesis is valuable reading for anyone interested in robot legs. It was exactly the envisioned vehicle, showing the limit of what brushless motor driven legs could do. With an unlimited budget, it could only go 4.47mph & it only lasted 20 minutes, so it would only go 1.4 miles. A reduced cost version would be even worse. It did show a robot with legs that only moved in 2 axes could still turn 360' in 6 seconds.
They provided some documentation on how the Minotaur worked: a 5.3Ah 4S battery, an STM32 realtime board, a blurry photo of motors which looked like t-motor U-10's. Unfortunately, you have to buy the robot to get the source code.
So a simple dummy mass with lever was built. The plan was to gradually increase power until the motor held it in 1 position. A human would trigger a fixed motion & manually turn it off. Hold a dummy mass in an equilibrium position was required before ever considering jumping movements.
It was assumed that since the motor previously pulsed 20V without overheating, with no load, any lower voltage could be held on for longer periods of time. At 16V, it held the load in the air, but in less than a second, the magic smoke came out. The multimeter showed no temperature change, but it still burned out. Minutes later, the multimeter rose to 40C. The only clue that it was heating up was it immediately started sagging after lifting the load.
There was probably more heating when it had a load. The test without load pulsed for only 260ms & probably missed the amount it was really heating up. The amount of time it can handle 20V would have been 200ms rather than seconds. Since holding a load like this for extended periods of time was required for any leg application, the use of direct drive motors was busted.
A better test would be to not rely on a human to turn it off, but make a minimum preprogrammed pulse just to observe the load moving at first, then slowly increase the preprogrammed holding duration to observe the load being held. This would allow the multimeter to more easily catch any temperature rise.
If anything, the leg would have to be really stubby. There is a lever diameter at which it could hold the load & jump. It would have to move more like a millepede than a dog & it wouldn't be fast enough or be able to overcome the required obstacles. The lever might have to be shorter than the motor diameter. It would purely be a science project.
The answer is yes. You can reprogram an ancient Hobbyking 18A as a stepper motor control. The mane advantage is a lot more current than your crummy L6234. The mane disadvantage is less precision. Unlike the advancement of radios & servos, there seems to be no limit to how long these ancient ESCs are going to be made, with the same ATMega8.
The problem is it requires software PWM. The FETs aren't connected to any hardware PWM pins, even if swapping the ATMega8 for an ATMega328. Your best route for higher precision is replacing the ATMega8 with a daughter board.
The precision was good enough for lion purposes. The Simonk firmware provided a starting point, but couldn't be translated directly from assembly. It uses only the timer 2 overflow interrupt. The output compare doesn't work for some reason if the prescaler is too low. Timer 2 is reloaded after updating every FET pin. It requires making a table of all the transition times in absolute time units, converting the absolute times to differences. An 8uS delay is needed between switching from P to N FETs. It happens in the interrupt handler, creating glitches when 2 transitions happen close together. Set the PWM to 2khz for maximum torque.
The UART pins are connected to FETs, so debugging requires bit banging a UART.
The key was the use of direct drive, brushless motors with stock windings to achieve high jumping. Jumping is the key to any robot running. It wasn't obvious in his thesis, but after extensively testing $75 gearboxes, he didn't use any gearboxes for the jumping. Pulsing the motors allowed the briefly high torque for jumping. More surprising was the holding torque they could provide.
They were the t-motor U10's at $330. The windings were only 95mOhm. A rewound motor for higher torque would have 14Ohms. The mane question is whether similar motors can be had for less money or if the robot can be scaled down for an after-tax budget. Maybe there could be a cheap motor for directional control & an expensive motor for the thrust. Maybe a servo could provide a degree of freedom.
The other note was how instead of 1 knee & 2 hip motors, he used 3 identical hip motors with freewheeling knee joints to achieve the same degrees of freedom. If it was passively stable in 1 direction, the leg could get by with 2 motors. To be passively stable in 1 direction, the CG has to be closer to the ground. That requires using jumping to overcome obstacles instead of leg clearance.
Unlike the quest for autonomous cars, virtual reality, & 3D TV's, there's going to be a single blog post someday about a commercially viable robot leg & it's instantly going to change the world. We've never had a commercially viable robot leg, so no machine has ever been able to get around terrain with the ease of an animal.
It used 250mAh in 87 minutes of driving. It stayed fastened, but took the room for a 2nd battery. Putting it in the back obstructs the sound. Putting the 2nd battery in the back would unbalance it. It might have to fasten to a side. What was convenient was having a removable battery which could be charged without taking out the speaker.
Congrats to the guy who got the last of the $10 Vivitars. That made the rest of us losers, even if $18 was equivalent to a single commute. Paying more than someone else for the same thing is still losing. $18 later, the 2nd bluetooth speaker arrived, with the goal of a robot speaker & job interview headset. It has become essential over the years for every vehicle to have sound & for it to be wireless.
It was quite loud & was probably light enough to use intact. The mane interest was hacking it to also take a microphone input. The innards revealed a 350mAh battery. 350mAh batteries have been revealed to last 30 minutes. It would have to be upgraded to support a full drive.
Most of it was a big, heavy speaker brick which can't be opened. It weighs 150g, which is the same as any other speaker. New speaker technology is needed if the same loudness is going to get any lighter.
Connectivity uses a very old BK8000L board. It uses a PN25F04 flash chip to store firmware. Audio amplification comes from an HT6871 3W class D mono amplifier. The positive stereo outputs of the BK8000L go through a pair of R + C's. The outputs of the 2 C's are tied together before feeding the amplifier's positive input.
An overly aggressive standby mode wipes out all transients, so it can't play game sound effects or metronome sounds unless something is always playing in the background.
Volume down is floating. Volume up is pulled to Vbat by a 10k, so the volume is...Continue Reading
So the bluetooth board caused a power supply noise problem when it used the TDA1517 as an amplifier. Must figure out how to defeat power supply noise, 1 of these days. Decided to discard the entire TDA1517 board & just use the bluetooth board with the old speakers because it would be the lightest.
The bluetooth board had some kind of series capacitor network feeding the amplifiers.
1uF + 20k + 100nF + 20k
Replaced it with just a
to try to make it louder. It now seemed just as loud as the TDA1517. The chinese datasheet showed a 58nF + 30k.
The BK3254 had no datasheet, but if holding down the power button to turn it on was annoying, it could be powered directly by applying 3.3V to pin 16. Unfortunately, it automatically reset this way so some other voltage was missing, perhaps 4.2V on pin 17. Never could find a 3.3V source on the board, so perhaps the chip implemented its own regulator & power button control.
Also, cable management nightmares proved a speaker is best powered by its own battery instead of the vehicle power.
It's probably best to use a bluetooth speaker as is, except for replacing the speakers & enclosure to make it lighter & louder.
In other news, discovered phones probably aren't cracking when the vehicles roll over but when they jump off curbs. When the vehicles always carried a shirt, the phones had cushioning & didn't crack. With a severe winter leaving the vehicles devoid of a shirt for many months, phones started cracking like crazy. They're also self destructing to get cheaper & thinner.
The lion kingdom still remembers when the most $10 would get were the cheapest, tiniest, most horrible sounding speakers which plugged directly into a walkman. The sound was completely unintelligible, but what a novelty in the old days to hear a walkman through any portable speaker.
There should be no disappointment in 25 years later, what would appear in Walmart but $10, Lipo powered, bluetooth speakers which actually sound as good as a boombox did. It's not just a case of something getting cheaper by getting smaller, but a lot more getting cheaper.
The purpose of this was not the speakers, but to gain a bluetooth receiver for a lot less than $parkfun's $25 bluetooth audio module. Managing the headphone cable had become quite a chore in 6 months of driving around with it. Sadly, the new speaker would be too heavy & quiet to use intact. It would require a much bigger robot to move around a decent speaker & loudness has taken precedence over quality.
The teardown was completed & the electrical signals probed.
It uses ANT8110 class D amplifiers which take 5.5V & output 3W. The only datasheet was purely chinese except for 5.5V, chipsourcetek, 3W, & class D. Though the reduction in voltage would be a huge convenience, it wouldn't be as loud as the 6W TDA1517.
The receiver is a Beken BK3254. Analog audio comes out of 2 pins at line level.
It requires 4.2V power at the battery terminals to power the ampilifier. The receiver can work on...Continue Reading
1) A robot dog which can go 10mph & traverse curbs.
2) A way of editing text on a tablet which is as fast as a keyboard + mouse.
3) Action cam with image stabilization
Quite different from years ago because the autonomous RC car is gone. That was deemed very impossible, for the moment.
1 is probably going to happen in the next 30 years. Thinking of ways to do it with standard servos. The mane problem with 10mph is the actuator needs to be more like a spring to generate a bounding movement. There could be a slow movement to charge the spring followed by a fast discharging movement. Traversing curbs would use fixed movements rather than adaptations based on sensor feedback. It would have to be passively stable.
The leading idea is standard servos for the knees & hips. Both rear hip servos would connect to a common, shmick servo. Only the standard servos would move for walking. For bounding, the shmick servo would fire simultaneously with all the rear servos to jump. If the servos are cascaded, the simultaneous motion of the 3 servos should produce enough velocity. The jump would lift off the front legs, then land on the front legs. The front legs would steer & allow the rear legs to reset.
This might move more like a rabbit than a dog, in which case the front legs would need another shmick servo, but the same general idea would provide any bounding movement.
2 is probably impossible. The Goog had some research which compared keyboards to...Continue Reading
So the Kyocera Hydro was a total failure. It was too slow to even type on, let alone load any web pages. The mane problem was the camera had permanently blurry regions. The lion kingdom didn't realize how important the phone cam had become until it didn't work. Japanese electronics once again came in far behind, subsidized by quantitative easing but not improving in 30 years.
In the world of phones, no matter what the problem is, the solution is to change providers. Boost Mobile had become cheaper than Virgin in the last 3 years. Virgin jacked up LG's bottom end, but Boost was still affordable.
The LG Tribute HD is a miracle for the price. It's lighter & thinner than all the other cheap phones. While noticeably slower than its predecessors, it's still faster than the Kyocera & a $50 screen of this resolution was unthinkable 3 years ago even if it's nowhere close to an iPhone. The camera isn't iPhone quality, but not a total failure like the Kyocera. LG cameras have always been just good enough. The internal flash was finally big enough to eliminate the need for an SD card. The Moto X wouldn't have been worth it.
It may have been inconsequential for someone who lived 30 years ago, but a screen where you can't see the pixels now seems essential. Part of the appeal may be that it looks like an iPhone.
Interestingly, below the Tribute HD are now $30 phones with decent screens, which can be used as real cheap standlone computers without a plan. A phone can be just a car music system or a dashcam for a lot less money than a stock device.
The day job has real nice phones which require hideous plans. The Galaxy S5 needs a $55 Verizon plan.
The 3rd LG Tribute disintegrated during a romp in the canyon. It lasted only a few weeks. Single women may have high medical bills, but their lack of exercise probably saves a lot on phones. Thus ended 3 years of Lucky Goldstar being the bottom end. They once adorned the Target shelves at $40. The Goldstars are now high end. The new low price leader is the $50 Kyocera at Target. For $10 less, there are Alcatels online only.
The low end reduced all their processors from 1.2Ghz to 1.1Ghz. They reduced all the cameras from 1920x1080 to 1280x720. It's 1 of those phases of economic stimulus where lower performance is rising in price to keep everyone employed. Mercifully, the Kyocera doesn't turn on when tapped. That bug definitely pushed the limits of millenial debugging ability.
Sadly, there's no way to get the contacts off the Tributes. Once their screens cracked, all the data was inaccessible.
This nugget caught the lion kingdom's attention last month. It seems to be doing the impossible: navigating a mountain dirt road with lots of shadows using only machine vision. For the 1st time in 11 years of UAV blog posts, there was no mention of GPS even being offered in the product. Of course, it didn't stay to the right like a running pacer has to do.
Knowing a convolutional neural network is just a storage & recall mechanism for images, it could have been done by training the neural network on many manual drives through the same 5 minute path. There's enough detail besides the path for the rover to ignore the shadows, determine where it is & steer to prerecorded headings. How the rover determined where it was in relation to the center of the path would be another trick. It could be a neural network trained on a bunch of offset camera views or it could be simple optical flow.
The key is the vertical detail from the trees & mountainside adding extra edges besides the sides of the path. It wouldn't do well on a bare path shaded by trees outside the camera view.
In other news, keeping with the modern slogan that the programming language is the computer, Tesla hired the guy who invented Swift & invented the LLVM as the VP of autopilot. 20 years ago, they would have hired a networking expert & 10 years ago, they would have hired a video expert. It would make sense for Tesla to invent its own programming language, since that's what everyone else is doing. Perhaps the world needs a programming language for neural networks instead of a library.
The LLVM has a good chance of replacing GCC. GCC has enjoyed a 3 decade reign, but ever since running it for the 1st time on an HP, it was slower than any commercial compiler.
There was a lot of hype about hybrid apps in 2014. Based on the internet, the verdict is there's absolutely no advantage to either method. The lion kingdom's experience:
GUI & speed requirements ended up beyond what the standard hybrid libraries could do without major effort. The browser incompatibilities & replicated effort when trying to optimize the interface for each phone were just as bad as native.
4 months of work on a webview implementation yielded nothing, with so much expertise being required in web development to achieve minimal results that the IOS & Android developers ended up going back to other tasks on their own platforms until another developer could be hired for the hybrid part.
It ended up requiring more people than if the hybrid part was done natively.
They're a dime a dozen, but still not sold commercially. Would feel sorry for anyone dumb enough to have to pay for one. The mane desire was to reuse foot pedals that were already around, but the Yamahas have an open circuit when they're active & closed circuit when they're inactive. The only solution was BJT's to invert the switches.
Making it work without batteries required 247k pullup resistors, so the pedals pulled down the bases when they were inactive & the camera bias voltage pulled up the bases when they were active. When the bases were pulled down, the resistors couldn't draw too much current to the point of triggering the shutter so they had to be large. It's an interesting use of a voltage to turn itself off.
they had to abandon supercooling at least part of the propellants. The helium will be warmed up to prevent pockets of frozen oxygen from forming. More of the engine exhaust will be bled off to pressurize the tank instead of using helium. They confirmed voids in the COPV laminations pooled liquid oxygen. When the liquid oxygen froze, it expanded & burst the laminations. More specifically, it was a void between the aluminum inner liner & carbon fiber overwrap. To save money, the overwrap was launched with voids.
You normally think of a tank being a barrier between stuff on the inside & stuff on the outside, but the overwrap on a COPV is actually saturated with stuff from the outside. The total propellant load thus consists of everything the engines need + a little bit for the COPV's to soak up. Presumably, the oxygen impregnating the tank can solidify as long as it doesn't form large pools.
Along with more expensive struts, they'll need super high tolerances for the overwrap, so explosion by explosion, it becomes less off the shelf & more precision specked like a shuttle. NASA spent years troubleshooting voids between solid propellants & booster walls. They avoided the explosions, but it took immense manufacturing standards.
They mentioned other changes besides warming the helium, but not what they were. They revealed the other changes involved spending a longer time loading propellants. They implied the helium loading procedure was changed during the fateful day from what it was in previous launches.
After 6 years on the wagon train taking heroic efforts to defeat noise, the preamp was redesigned again. The discrete transistor amplifiers are now an LF353. The discrete transistor was nifty, but impractical for fully balancing 2 channels. The LF353 subtracts the differential inputs & amplifies the result with minimal components. This eliminated the noise completely & finally allows the Zoom to record 4 channels. It also seems capable of sharing the power supply with the Zoom without any noise, but it's not soldered for it.
The mane problem is the pots being too close together. The mane cost in going to 2 balanced channels is 2 stereo pots & 8 wires to connect them. Previous work with digital pots ran into a lot of noise from the SPI & a display to show the current level. The balanced inputs may defeat the noise, but the analog controls have proven simpler.
The LF353 circuit had to deviate from the goog slightly. The 1k's had to be fixed. Only the 100k's could vary. It needed 1k's on the outputs to isolate it from Zoom noise. Spudger diodes prevent the 48V pop from blowing it up. The virtual ground is 5V. It needs 12V to be happy.
All support for 2V microphones & dynamic microphones is gone. It's now just a phantom powered 48V preamp.
The moral of the story is any microphone is going to need a balanced amplifier in its 1st stage, right down to your cell phone mems microphone. The balanced signal needs to be retained all the way until the line level or power supply noise will always be a problem.