Jack Crossfire's blog View Details
Posted by Jack Crossfire | Jun 01, 2016 @ 12:28 PM | 2,283 Views
For all the hype in the last 11 years, SOMA has never been known for the Autonomous Car Startup. Except for 1 other ACS which disappeared years ago, it's always been about the Classic Social Network Startup, which made this interview all the more fascinating.

We go riding with George Hotz and his $1,000 autonomous car (13 min 51 sec)

Once past the usual tired antics of the pitch, he dropped a few nuggets not in any other interview. His algorithm is based manely on machine vision, probably another convnet algorithm. Segments showed a front radar sensor & an external GPS antenna. It failed to detect a car in front of it. It started tracking an exit ramp, requiring him to take over. He claimed it detected a Botts dot lane marking which no other algorithm could track, but more realistically it detected the tire wear.

As usual, it's a very early prototype which can't change lanes. As with personal experience, he's trying to train the algorithm with videos of previous drives. Also matching personal experience, it didn't work at dusk. Either the sky turned white & the road turned black or it couldn't handle the shadows. He didn't say which was the case. He did say it worked at night, showing edge detection was probably 1 of the convnet stages.

Neither is there any information about how well it corrects a vehicle significantly out of the lane. Unless the training set involves a lot of drives off the road & in every possible offset from the lane markings, supervised learning algorithms are manely good at staying in the lane if they're already in the lane but bad at correcting if they go off.

Supervised learning not only requires many videos of driving, but a human must manually enter in where the lanes are. There's still no bulletproof solution to shadows, making any effort in manual training a waste of time.
Posted by Jack Crossfire | May 29, 2016 @ 10:33 PM | 2,532 Views
In the scheme of things, microcontrollers are brought up in 4 steps:

user program

Another 2 weeks got all the Feiyu peripherals working. It uses the SPI, I2C, PWM, ADC, a quite massive number of peripherals. The mane issue was the I2C driver. It doesn't work if a UART sharing the same pins was enabled at any time in its past. The UART has to remane disabled from startup. So a new bootloader which didn't enable UART3 went in the board which uses I2C. Ideally, all 3 bootloaders would only enable UART3 if the user sent it a passthrough command, but for now, 1 bootloader is hard coded with it off.

Couldn't get I2C to the 1Mhz probed. It only went to 100khz, which was fast enough for over 400 IMU readouts/sec, but not as low as Mr. Feiyu himself in latency.

The mane clock had to be reset in the user program, after the bootloader already set it. The C library start functions reset it back to 8Mhz.

SPI had a strange bug where the SPI_FLAG_BSY bit had to be tested instead of the SPI_FLAG_RXNE bit. The SPI_FLAG_RXNE bit cleared long before it was done transmitting. Also, the user has to read DR after every transaction or it won't fire again.

The hall effect sensor produced valid data, despite the lack of any datasheet.

The mane trick in enabling the motor was creating deadband between the N & P pulses. That required setting the TIM_BDTR_OSSI in the BDTR register & setting the DTG bits to the deadband time. There are...Continue Reading
Posted by Jack Crossfire | May 28, 2016 @ 09:36 PM | 2,224 Views
Deuce Bigelow isn't known for before & after photos, because his modules aren't really as inflatable as balloons. The 1st 2 were privately funded cylinders launched on Proton rockets, back when Proton rockets worked. They didn't expand in length & only slightly in diameter.

10 years & a few waves of RIFs later, they finally launched their last module using NASA funding. 2 months later, it was inflated. There are no official before & after photos, but some careful editing revealed the size change. It was a bit more impressive in length than the 1st 2, but still not inflating significantly in diameter. It has no life support, relying entirely on the space station for air. It can't be used for living space.

Their plan is in 5 or 10 years, with any luck, getting a 43,000lb module up. Too heavy for a Falcon 9, so it would require an Atlas 5. There are no diagrams of how an inflatable module is built, but it is known that the walls contain bladders which dictate the shape of the module. The bladders are filled to a different pressure than the habitable space. The walls of the NASA module were not disclosed. The walls of the hypothetical 43,000lb module would be 18in thick.

For all the efficiency of spherical space stations, the modules can't be spheres. They have to be confined to the shape of the payload fairing, which is still a cylinder.

For all the complexity of inflation, inflatable modules don't increase the amount of junk you can put in...Continue Reading
Posted by Jack Crossfire | May 28, 2016 @ 12:17 AM | 2,309 Views
Not easy to think about anything else, after the video. You know the one. We all grew up watching the Apollo 4 stage separations & the Apollo 6 earth departure stage departing while the camera stayed behind in the S-II stage, & the shuttle departing the boosters. Nothing was quite like watching a 1st stage manipulating grid fins & engine burns to return from a parabolic arc to a powered landing on a tiny X in the sea.
Posted by Jack Crossfire | May 22, 2016 @ 09:14 PM | 2,272 Views
The answer is no. Elon Musk wasn't at the Tesla factory 5k.

Started getting nauseous near the end, but overcame it by throttling back. Real hot in the factory. 1 mile of it was in the factory, winding up & down what seemed to be the entire assembly line, though the fake test pilot couldn't differentiate different car models or different states of completion. They all looked nearly finished. The robots were quite enormous & organized into discrete work cells, separate from areas with manual assembly. They had stopped production, but some people were walking around with parts, not happy to be working on Sunday.

A lot of people walked the course, making a factory tour out of it. The fake test pilot sped through, as usual. Just can't slow down in a 5k. The magic would have been if the robots were left on, revealing production techniques. Surprised fake test pilots could sustain such a high speed for so long. Threw down an average speed of 7m29s/mile.
Posted by Jack Crossfire | May 21, 2016 @ 01:44 AM | 3,123 Views

It wasn't that Neistat was busted for riding his Boosted board on a bike path. It was that he was busted 1 week after the fake test pilot was busted for driving the lunchbox on a bike path. Government budgets must have finally grown enough to enforce those things.

It was a routine 18 miler just after dark. It was completely empty. The vehicle & human had sustained 9min/mile quite easily for 3 miles. Saw 2 cop cars parked next to Walmart. Saw some years ago, so it seemed to be a common beat to park next to Walmart & eat the donuts. The cars were empty. Passed a homeless guy who was standing in the middle of the trail, seemingly confused. In the distance were 2 guys, probably dog walkers.

The 2 guys shined flashlights at the fake test pilot, not uncommon. What was uncommon was the lights followed me as I got closer. Saw they were cops. Then they said "Hey! Can we talk?" It was the 1st traffic stop of a runner for using a robotic pacer, probably in human history.

The elevated heart rate & breathing from running 3 miles made for an intense dialogue. "What is this machine doing! You know motor vehicles aren't allowed! Neighbors & bikers complained about RC cars! There's a $250 fine. We're not going to fine you now, but someone else may not give you the same courtesy!"

As the fake test pilot cooled down, they got more peaceful & asked how fast it went, if I...Continue Reading
Posted by Jack Crossfire | May 17, 2016 @ 11:27 PM | 1,989 Views
Basically, a loudspeaker for the exercise robots would be nice, but dynamic speakers are too heavy to cost effectively move. Thus, a practical use for electrostatic speakers appeared. Electrostatic speakers are used in a Best Buy demo for a game. Typical for the age, there's very little online information on how to make an electrostatic speaker. There are no specific voltages or currents. The only fact is the driver part of an electrostatic speaker is much lighter than a dynamic speaker.

The only useful parts list is on:


There are no useful schematics, but notes on the speaker performance here:


The Elliott speaker uses the audio amplifier to power all 3 plates. The Jazzman speaker uses a separate supply for the diaphragm. They all passively power the stators by stepping up the audio amplifier voltage.

Presumably, the stators go at 1000V. The diaphragm goes at 2500V. The required power is unknown, though they say the speaker presents a 4uF load which requires a lot of power, maybe even more power than a bug zapper or cold cathode. The transformers are very large, but what if the audio amplifier could be actively stepped up by a semiconductor solution. It's unlikely a semiconductor solution would generate enough power, but it might work for a very small speaker.

Barring the electrostatic speaker, there's using some combination of piezo buzzers to obtain lower frequencies. Perhaps 2 piezo buzzers on opposite ends of a large resonator could alternate single pulses.

10 years ago, there was an effort to make a speaker which used any plastic surface as its resonator. It was targeted at eliminating speaker grilles in gadgets by using the enclosure itself. It went absolutely nowhere.
Posted by Jack Crossfire | May 15, 2016 @ 07:49 PM | 2,065 Views
Recovered Falcon 9 fotos are hard to come by, due to the secrecy of a commercial operation & the lack of interest compared to more important matters like Trump's imminent non election.

US Launch Report is the only guy covering recovered booster processing.


Like the shuttle booster recovery, every flight is still a major construction project. Still kind of novel seeing ordinary construction cranes of the past 30 years handling the alien looking landing legs, like humans with limited means trying to figure out how to move parts of a giant alien robot.

In the 1980's the visionaries we read about were Werner Von Braun, Thomas Edison, Nicola Tesla. The common theme was they all existed in the past. There weren't any visionaries in the present because the odds were too remote for our generation to coincide with someone like that. Musk is the closest to being a living Werner Von Braun, but Werner Von Braun actually saw his giant moon rocket come to fruition, even if his Mars vision never happened.

There isn't much to the Musk story which hasn't been done before. Kerosine rockets & labor intensive recovery of damaged boosters have been done before. The talk of a mass produced electric car, gigafactory & self driving cars is still talk. He hasn't realized the vision of airline style rocket reuse or travel to Mars. His car & solar companies are financial disasters.

At least there's...Continue Reading
Posted by Jack Crossfire | May 14, 2016 @ 08:54 PM | 2,350 Views
After many traps for young players, the Feiyu began running a custom bootloader. Key points were converting the GPIOs from JTAG, changing a bunch of HAL code . Once complete, there will be no need for the Nordic board or all the wires. The trick is going to be daisychaining all 3 boards & fitting everything in flash. It may seem pointless to reconstitute such a low end chip, but it is what made the Feiyu the small size it is. The mane problem is all this is for just 1 copy of 1 product that was only made for 3 months.
Posted by Jack Crossfire | May 08, 2016 @ 04:23 PM | 2,199 Views
So the mane processor is an STM32F103C8 64K flash, 20K RAM, 72 Mhz, 48 pins. It breaks out single wire debug on pins 34, 37 for raw programming. It also breaks out 3.3V, GND to the programming header.

There's no distinguishing feature for the chips to detect which board they're on. It must be programmed in the bootloader or some initialization handshake.

MOSFETS are connected directly to GND & battery V+. They're some kind of motor controller packages with complementary MOSFETS.

switching regulator reduces battery to 5V
linear regulator reduces 5V to 3.3V
STM32 VDD comes from linear regulator
LMV358 VDD comes from linear regulator
hall effect VDD comes from linear regulator

Header entering yaw board:

RX/yaw PWM -> 100R to pin 43/UART1_RX/I2C1_SDA/TIM4_CH1
TX/pitch PWM -> 100R to pin 42/UART1_TX/I2C1_SCL/TIM4_CH2
Mode PWM -> directly to pin 41/TIM3_CH2

Header exiting yaw board:

100R to pin 21/UART3_TX/I2C2_SCL/TIM2_CH3
100R to pin 22/UART3_RX/I2C2_SDA/TIM2_CH4

Board to board wires are straight through. The boards communicate via UART at 2Mhz. Colors are reversed on the IMU wire but I2C connects to the MPU-6050 as expected. I2C goes at 1Mhz.

Hall effect sensor uses SPI. Clock speed is 561khz. Chip select goes low 1st. The STM32 sends a command in 16 bits, switches the SPI to input & reads the result in 16 bits. Chip select goes high last. Sequence repeats every 300us. Doesn't have an initialization...Continue Reading
Posted by Jack Crossfire | Apr 30, 2016 @ 06:10 PM | 2,993 Views
Future of Faster | The PUMA BeatBot (1 min 30 sec)


It only took 3 years, but MIT & NASA finally caught up with the blog. Their version of the pacing robot uses line following for navigation & they didn't seem to have solved the manual control problem, yet. The line following is based on IR leds with a video camera so it only works at night, on a track.

It won't be long until they discover line following is not worth as much as an ergonomic manual control.
Posted by Jack Crossfire | Apr 27, 2016 @ 12:43 PM | 2,427 Views
After 2 weeks of round the clock commuting & moving, the home network came up again. It was sheer brilliance for Comca$t to make every one of their rental modems a public hotspot & it's amazing AT&T never copied the idea. When the idea 1st came out, the internet screamed bloody murder. What would be the impact on performance for subscribers when their rented modem was shared with a user on a different account? Of course, it was more AT&T screaming bloody murder & the internet just doing what it was told.

In the fullness of time, the outcries disappeared like any dead celebrity meme & the sheer brilliance of the move arose. Despite the insane cost, most people still rent modems instead of buying their own for half the price. Every one of those millions of modems automatically pops up on every wifi enabled gadget, a gigantic window saying "XFinity". It's the biggest customer sponsored advertising blitz, ever.

It takes very forward thinking to get to an advertising blitz this extreme. It began when the wifi standards were ratified. Someone had to realize allowing a router to push a web page for logging in was a useful feature. The true power was in the standards process.

So when finally subscribing to Comca$t, after years of endless XFinity popups on every wifi device, the hotspots instantly worked. There was no need to buy a modem at all or wait for an installation kit in the mail. You could just use your neighbor's wifi & it even went up to 8 megabits. Not sure if the speed was throttled based on the account. Every modem automatically signed in.

When the wired modem finally arrived, it required jumping through an activation hoop. After this, the wifi hotspots no longer worked. It seems they only allow use of the hotspots before activating your modem. Then, they require upgrading your plan, yet another marketing move based on human fear of denial.
Posted by Jack Crossfire | Apr 22, 2016 @ 05:25 PM | 2,517 Views
There have been so many man rated quad copters, EDF powered hoverboards, mini helicopters, that only go 5 minutes, can't get above ground effect, or never got finished, the Flyboard looked like more of the same when in fact it was different. The key is to consider too this day, the smallest flying machine that actually works is the original hydrogen peroxide jet pack. It goes 30 seconds.

The Flyboard is the 1st to be at least as compact as the jet pack, but go much longer than 30 seconds. It uses 4 hobby size jet engines at $10,000 each. The fuel is in a back pack with a hose going down the pilot's leg. If the fuel line gets kinked, he loses $40,000 & dies.

The remote control only controls thrust. The pilot leans to control direction. It probably has foot switches, but his feet are rigidly mounted. There could be a trick involved in rigid foot mounting, yet allowing enough motion to provide user input.

Hobby jets probably have enough response to stabilize it by varying engine thrust. There's no thrust vectoring or reaction control thrusters. 2 red EDFs on the sides counter the torque of the jets & provide yaw.

There's nothing in it which couldn't have been done 10 years ago, except for coming up with the money. Borrowing $40,000 was once equivalent to borrowing $400,000 today.
Posted by Jack Crossfire | Apr 19, 2016 @ 11:04 PM | 2,850 Views
To not feel so bummed about the home made copters of years ago being stressful & difficult to keep out of harm's way, it's time to start compiling Neistat's quad copter crashes & tribulations.

boat crash:


loss of signal:


bridge crash:


So basically, despite years of multiplying valuations & redundancies, even the most foolproof quad copter still isn't much more foolproof than the old days. The attempts at obstacle avoidance by sonar & IR obviously haven't translated into anything meaningful. Neistat explained that he gets away with flying by having 4 quad copters & chewing them up.


He explained that he was getting all the selfie shots by semi autonomous control, not by any dedicated feature:


Ergonomics were probably the biggest factor in doing it manually. Given the amount of flight time & length of the final shot he gets, it's probably too cumbersome to set up a "cable cam" , waypoints, "leash cam", or geofence.

Still suprising he has to carry all that gear. There would be no way he could use a 3DR copter because they're too big. The Phantom is just small enough.

Ground vehicles have no worries. They can precisely navigate the city streets. Combining their single handed controller with a quad copter would be the ultimate selfie system, 10 minutes at a time.

The videos of the Phantom 4 at low altitude do look rock solid precise. It uses sonar for altitude & dual cameras for optical flow, but also fuses that with the accelerometer.
Posted by Jack Crossfire | Apr 13, 2016 @ 08:54 PM | 3,021 Views
The amount of soot appears related to the amount of time the surface spent at ambient temperature, so the gradations in soot can show the amount of LOX left in the tank during each re-entry burn. 33% of the total LOX capacity was required for recovery. Of the 33%, finer gradations show the 1st & 2nd re-entry burns with 15% left over for the landing burn. The line dividing black from white would be the edge of the common bulkhead, since there is no intertank section to provide a smooth gradient.
Posted by Jack Crossfire | Apr 09, 2016 @ 08:14 PM | 2,497 Views
Managed to get the USB->serial dongle working in VirtualBox, since it's a ttyUSB instead of a ttyACM. The problem is the gimbal has to be turned on after the firmware update tool is already listening. The firmware update tool waits for a start code so it can force it into bootloader mode. PWM did indeed have to be off. TX came from the pitch PWM. RX went into the yaw PWM.

So the Feiyu turned out to already have 1.14, the latest firmware. That left waiting for another firmware update in the future or programming it from scratch. Feiyu already discontinued the mini 3D, so another update is unlikely.

Tearing it down revealed a bewildering nest of miniaturized boards which would have to be reverse engineered. The same board is replicated for each motor. There is only 1 IMU.

Next, we have an MV358 op amp, given away by some RC pairs next to it. It probably amplifies a signal from a chip marked only 01281000, undoubtedly a hall effect sensor despite the lack of any Goog results. It's directly under a magnet on each motor which serves no purpose other than giving the motor's orientation. These magnets are circular polarized to give a varying field for each position.

The same firmware probably runs on each board, relying on the placement of jumper resistors to know which board it is. Communication between the boards, how to change the firmware with the existing bootloader, how to control the motors are all the great tasks. It would have been a lot easier if they just allowed configuring the product.
Posted by Jack Crossfire | Apr 09, 2016 @ 02:45 AM | 2,783 Views

Another small step forward. Landing at sea increases the amount of mass which can be launched on a reusable rocket. Still up to the voters & government to determine when humans will finally be able ride it. Their 1st launch of a human is entirely dependant on your continued support of the CCDEV program. The human mission to low earth orbit would be the easiest to recover the booster from, the baseline mission for the reusable system.

You would think there would be competition between customers to be the 1st to pay for a used rocket. It would be quite a publicity boost. The 1st is going to be the one in Wikipedia for all time or at least as long as it keeps raising donations.
Posted by Jack Crossfire | Apr 07, 2016 @ 03:05 PM | 2,545 Views
Current usage with camera recording at 8.4V:

0.85A with motors not deflected
2.3A with all 3 motors fully deflected

The old pro doesn't work without battery. There's no point in using the power supply umbilical unless you're previewing video. Even then, the power lead should be desoldered. It's easier to use the pro battery in its included spot than carry around an extra gimbal battery.

A dummy battery could go in the pro that would trick it into working without a battery. It probably just needs to see 4.2V on a pin, but no spare battery exists to sacrifice.

Current usage without powering camera at 8.4V:

0.25A with motors not deflected
1A with all 3 motors fully deflected

When powering the camera, a converter reduces current usage for higher voltages. Without powering the camera, the current rises slightly for higher voltages, but with no increase in torque. Minimum current is always 0.25A. The torque seems just high enough, but could go higher without burning out the motors.

Motor control is quite a bit more robust than past gimbals. There's no skipping or studdering. It's like it has encoders giving it motor position, but it doesn't. It's using gyros attached to all 3 motors to gleam phase in addition to orientation. It seems to have an algorithm change which invokes the gyro based encoding when the angles get offset beyond a certain amount. This allows it to recenter instead of studdering.

It was deemed good enough to hard code the mode & speed in...Continue Reading
Posted by Jack Crossfire | Apr 04, 2016 @ 12:39 PM | 1,997 Views
The 1st assembly of the Feiyu was disappointing. It wasn't as transportable as hoped, but it was more compact than keeping the shock absorbers. Wires hung out everywhere, but the assembly can probably be wrapped. The aluminum arms are going to bend over time. The bolts need reinforcing tubes. There's not enough clearance from the handle to easily turn an allen wrench. The pins for calibrating the horizon need to be flush on the handle.

The power switch managed to stick on by reflowing the hot glue after gluing it on. Reflowing hot glue by blasting it with the heat gun after initially tacking it has been effective at sticking difficult parts.

The power switch was deemed necessary because the Feiyu has to be powered on in the orientation it's going to stabilize in, but it's a fragile arrangement with lots of wires. A quick test should reveal if the power switch is necessary or if the battery plug, a MOSFET with delay, a tact button, or app button for enabling the Feiyu can do the job.

It was decided to remove the 2 pots for mode & speed. Those are better controlled over bluetooth app. The voltage regulator needs to be replaced by a 3.3V & bluetooth connected to the debugging port. Need to measure the power consumption of bluetooth.