A day later, a new motor plate & sensor plate arrived. This one had more room below the sensor plate. The long awaited motor & sensor rebuild happened. Nothing obviously moved since the last rebuild. There was no corrosion.

The sensor has bee dick margins. Maybe the shaft could bump into it. It took a full rebuild to discover there's still very little side to side margin. That margin is only visible from above. That would take moving the standoffs closer & rotating the hall effect sensors 45 deg. That might be the next step.

------------------------------------------------------
8 miles with no issues, so team sensor plate spacing gets a win.

Walnut Creek run timelapse (6 min 4 sec)

It was adventurous. There were many right motor failures for the 1st 5 miles, then it seemed to stabilize. After yesterday, this ruled out stiffness of the shock absorber & pointed back at the farsteners moving. In the worst case, there could be a contract manufactured motor plate out of aluminum. From the very limited search results, it could be $400. Most animals only want 2 axis milling or circuit boards.

The leash had another software malfunction where it started driving too fast instead of too slow, after a motor calibration. The mane confuser had to be reset rather than the leash. 1 theory is the sync codes getting out of phase.

Besides those problems, the leash was amazingly stable. Never saw it track with such low oscillation. The trick was just a low P & a very low D.

After a day of calibrating the motor every mile, another round of testing began. Man handling the motor between calibrations didn't shift it enough to change the table. Both motors showed similar changes in calibration tables despite man handling.

Another idea emerged where the problem wasn't the motor shifting around on its mounting plate, but the shock absorber being too loose, causing the sensor module to bump against the enclosure. There's not much clearance in there.

The long term plan is to make a smaller sensor module with more clearance. It's going to be a major rebuild. The shock absorber was tightened in the mean time. Hacked the motor calibration software to make it go faster & hopefully generate more accurate tables.

It's having a better phase.

The ankle was good enough after 2 months to go 10mph again. It seems walking at a 14-15 minute pace does more recovery than slow walking. Also, carrying extra weights might go further.

There is evidence that a new battery influences speed most of all but lions wouldn't discount the nose cone. The folding nose design together with the latest handle design make it a very streamlined process with no tools, to go from a forward facing nose cone to a forward facing leash. This was quite a difficult design to arrive at. Maybe the higher precision ABS manufacturing in China could create something more efficient.

It's been at least 17 years since lions began playing with gyros, starting with the mighty Invensense IDG300. Noted as the MPU6050 has aged, it's shown more drift. It might be because of age or it might be just many accumulated years of experience in different conditions. Today, it started at -40, drifted to 163, & finished back at -40. This coincided with it facing away from the sun for an extended time, facing towards the sun for an extended time, then carrying frozen food. Obviously, it's responding to temperature change. It would have been undriveable without constant resetting. It might be more sensitive to temperature than it was 10 years ago.

The temperature sensitivity has proven more extreme at low speeds because it spends more time in shade or sunlight. When moving at high speeds, it equalizes at an average of all shade & sunlight. Figured chinese gadgets use a temperature table to calibrate their gyros but this was never precise enough in real experience. They might also update the temperature table with more precise readings when the gadget is stationary. It might be worth revisiting the temperature table to overcome the need to constantly reset.

F35B Stabilized (7 min 32 sec)

Finally got some footage of the hover. As much as the internet hates the STOVL capability, a supersonic stealth aircraft is a huge capability to bring to a small boat. Without STOVL, it would need an aircraft carrier. It's remarkable to lions that they do hovering operations every day without any engine failures & crashes. At almost 1000 produced, it's a much more capable airplane than the internet lets on.

https://www.theverge.com/2023/8/10/2...-enterprise-x2



Another one bows to the Chinese juggernaut, so you can't blame open sourcing the autopilot this time.

DJI Mini 4 Pro vs The Gauntlet (Autonomous Tracking Test) (17 min 35 sec)

Reviewers found DJI's obstacle avoidance to be inferior but cheaper. The big features in the last 10 years were tracking & obstacle avoidance, yet to this day no-one outside China really knows how machine vision trackers differentiate a person in a crowd. There's some evidence they might use chroma keying on top of a simple person detector. Person detectors are pretty germane nowadays. You can train efficientdet to detect just humans. Chroma keying is still subject to the vagaries of white balance & lighting. A head recognizer would be a game changer. There's no known head tracker which can recognize a head from all angles. There are only face trackers which only work from in front.

Since lions are using a pose detector, the idea occurred of creating a separate color histogram for each body part & using the histograms of all the body parts to recognize a lion in a crowd. Multiple body parts could enhance the detection. There isn't going to be anything more reliable than color histograms & identification of any kind might just require wearing differentiating colors, ideal lighting.

When only 1 animal is visible, it could constantly update its histogram table, but there were cases in testing where someone other than a lion was the only subject.

Testing it is the real problem. Lions just aren't around crowds except very rare cases.

After getting good results from stretching the 1st 4 layers on PLA & TPU, there was still a risk of a nozzle jam when printing the remaneing layers of a tire. The latest idea was to stretch all the layers in the tire to .36mm after slicing .32mm. This requires reducing the TOTAL_H in the helix.c program from 40 to 36mm for the rear tires. There's 1 extra mm to account for Z compression. The mane unknown is layer adhesion after stretching.

The front tires are harder because their CAD model defines the height. The best solution might be stretching the wheels since they're not limited by a motor size. The result of this was to make the front tires softer but wider. The front tires never had nozzle clogs, so it was all a bit unnecessary.

For the rear tires, the layer stretching still didn't work. The 1st 2 layers clogged in their inner arcs, the same as before. The 3rd & 4th layers barely made it. The remaneing layers extruded easily.

The best strategy is still listening for the long extrusion & watching it until it succeeds. Would say it's over extruding by a large amount for the rear tires. Instead of stretching all the layers, it might require the 1st 4 layers to be .40 & the rest .32. It would be nice if it conditionally paused before the long extrusions & beeped but there aren't enough pins.

It could be the polygon ordering. The front tires start each layer with the inner arc. The rear tires end each layer with the inner arc. The...Continue Reading

https://youtu.be/pJlQqTHXtws

Some speed tests showed the full jetson tracking system could hit 6:50 when set to 6:30, without a nose cone & with a single battery. It's maximum speed was around 6:30 depending on wind & terrain. Quite a bit faster than the raspberry pi was, because of improved balance & suspension.

The best 3rd person shot is a static shot pointing downrange. Manually panning at the lowest speed works by luck. There could be a hack where it does a fixed speed for a fixed time but this would be harder to add the effect to. There's a chance of it losing the signal & not getting the correct time. The panning does show more lion before the effect would appear.

There's a chance the tracker could work in 3rd person but the tracking camera would need to point horizontally. More likely, it would track the wrong animal.


Looks like the tracker lost the lion in a few profile positions with good lighting. It's 1 point in favor of making it write out a vijeo file.

The longer it goes without problems, the bigger the problems when it finally falls over. CA apparently doesn't bond to glossy PLA so a button cover just popped off. The button covers can be replaced without a full tear down, but it requires no debris to be anywhere. In this case, debris required a full teardown & removal of the button to clean the debris & get it to seat properly. This revealed a full load of salt. It was like the waterproofing measures were never even there.


This salt didn't come until the peak of summer, when we had abnormally high humidity. Helas, it also didn't come until after creating a hole in the bottom of the sock & removing the body tape. The idea was for water to drain out, but there's a chance it really drew water in by surface tension. Without the hole & with body tape, water was accumulating on the flat parts but not getting in. With the hole & without body tape, water was accumulating on the sides & getting in. Body tape didn't allow any routine cleaning & was a pain to replace. It might have redrected the water from the sides to the flat parts.

The only thing to do is to reintroduce the body tape & then try combinations of a hole less sock & body tape.

There's growing interest in having this controller auto center every time it turns on. It already has required a manual centering every time it turns on & several manual centerings in the course of a run.

Motor angle resolving was never a sure thing & the right motor became unreliable after the last rebuild. The algorithm uses 1 table (hall0_table_ptr) to convert a raw hall effect sensor voltage to an index value. It uses another table (motor0_table_ptr) to convert 2 hall effect sensor index values to an angle. The sensor voltages are prone to DC offset. The bad motor has a lot more DC offset between calibrations than the good motor but even the good motor has DC offset. It would be better to detect a ratio of sensor voltages. Mechanical alignment has never been good enough to rely on simple trig functions. There's a lot of play in the different bits. It's not clear why some rebuilds have very stable alignments & some rebuilds are very unstable.

There are off the shelf angle sensors with the mechanical accuracy required for trig based decoding & which don't need any calibration. Commutation just happens too fast to use I2C. It may just be time to go back to a conventional 3 sensor system. Commercial hobby motors have the same resolver magnet separate from the motor armature but with 3 sensors. That would supposedly eliminate the calibration step.

-------------------------------------------------------------

Merely wiggling the motor in its bolt mounts can get the same voltage change. There's a lot of play between the bolts & the flange. Going to M3 instead of 4-40 would make it a hair tighter but the holes have worn enough over the years to...Continue Reading

The latest thinking was to bolt the jetson on the truck. If it gets run over, it's a $150 paperweight & lions won't be inclined to burn $500 on a higher end model. Electronicals aren't getting cheaper anymore.. #inflationtargeting


There's still a chance of letting it flop around in a padded enclosure. It's a lot bigger than the raspberry pi. Nothing is stopping a string from reinforcing the handle & it might be necessary for the 'pro.

The 1st test was for tracking robustness without the 'pro. The limitations of USB wifi & phone app restrictions make screencaps no longer a viable way of debugging the tracker. There are definitely problems with task switching the app & frames getting split.

Still prone to detecting trees like efficientdet. Generally less false positives & more robust than efficientdet. Had no cases of it actually chasing the wrong subject like efficientdet had. It suffers from false negatives the same way face detection did. Being able to scan a full frame at a time is definitely helping.

Nothing is as good as a custom trained efficientdet at 21fps, but not being able to run a custom trained model on the jetson & not having a head recognizer make body_25 a viable alternative.

Since the webcam got lighter & we're not tracking heads, there is a desire to move it higher up again & angle it lower. Past experience showed body_25 does better when the head is visible though. There might be a more flexible USB cable. The camera should always be in front of the handle now, since the confuser is always in back. All that iron & computing brought the power up to 400mAh/mile.

After much CAD modeling, filament drying, printing, & loathing, both tail lights got another refresh. The heat sink, sil pad, led stack got a dose of glue stick. It's unclear if the glue stick did anything. The mane problem is the parts have to move around after assembly, so adhesive tape wouldn't work. The best tape system might involve a jig where the LEDs go down 1st, then the sil pad goes down, then the heat sink goes down. With the current filament condition & the holes, very little light should be getting wasted. The current system is still the most compact coming to mind.




The leash began enjoying an alternative use as a clamp. Ideally, it could also be a strap. With the handle in place, it could be a strap. Without the handle, the trick is going to be getting the clamp through the holes.



Good news: the latest container design is rigid enough to not require a string to keep the handle in.
Bad news: the latest container is too rigid to let the handle out without changing the shape of the handle. A transforming handle was devised which can slide out by only bending the container 5mm. As the PLA ages & the container loses its resistance, this could be adjusted or it could go back to string.

Without string, the leash could be used to strap a payload on. The latest containers are big enough & lion diet has changed enough to not need to do this. The container is rarely filled all the way anymore.


-----------------------...Continue Reading

Ended today's drive after 1 mile, on a bumpy road. Given the recent cable problems, it's an automatic cable replacement. It was a transient failure & couldn't be reproduced on the bench. Elected not to reboot it in the field but drove it home in leash mode, just in case it wouldn't reboot & become a brick. IMU failures remane extremely rare. Easy to forget not long ago, the only way to drive an RC truck was RC because leashes only existed after last year. Leash performance only reached its current usability in the last month.


In the tail light department, the leading contender is adhesive thermal tape.

https://www.amazon.com/AIYUNNI-Condu.../dp/B08611W9FN

Double sided thermal tape is pretty commonly used for attaching heat sinks where there's no place for farsteners. It would keep the tail light sandwich in place & improve the cooling. The question is with free shipping now requiring $35 of orders, is amazon prime worth it.

The problem with this is the parts have to be shifted around after assembly. Another idea is good old glue sticking.

The 1st wear related repairs since June 20 happened. The mane power cable had a clean break despite being a 20 gauge silicone flex cable. This of course took many hours of replacing connectors & wires to find. It got replaced by a long desired doubling of capacity. Every intermittent electrical failure automatically involves tearing down & rebuilding an entire circuit.


The accessory cable is chewed up from many car accidents. It'll need replacement next. There is evidence of the wire passages through PLA causing wear. Those passages need to be rounded.

Finally, the tail light stack isn't staying together as much as hoped. The next step is adhesive.

https://twitter.com/greentheonly/sta...40665330515969

https://insideevs.com/news/671961/te...dar-confirmed/

Elon quietly reintroduced front facing radar. It's obviously to address phantom braking, but it makes lions wonder if the vision wave is suddenly going to end & all the startups are going to switch to high resolution radar next.

Lions gave up on self driving after a few tests with Sped up robust features & home made vanishing point detectors. Affordable cameras were showing their limitations but everyone else was sticking to vision on Elon's orders. Very high resolution radar could be a different story.

An autonomous 1:10 scale truck which follows a lion, complete with terrain mapping & obstacle avoidance, remanes a thing lions don't completely discount will ever be achieved by a large group with a lot of money, but it's not likely in the next 100 years. It's a harder problem than what food delivery robots do since it doesn't have prior knowledge of the route or the terrain.

The best outcome might be incrementally more partial autonomy than it already has. A path following mode with obstacle avoidance seems attainable with radar. The operator would point it in the right direction & flip a switch on the truck to stay on the current path. This application still feels like it's heading towards GPS or some prior knowledge of the route. A phone app would select the destination point for computing heading while the existing leash would...Continue Reading

The latest version came from very well dried filament in the latest solar drier. This one is just a webbing keeping the LEDs in. It projects a lot more light on the room but appears as a smaller pin point. Print time & amount of material was vastly lower. The headlights now feel lacking compared to the tail lights. Lions are more at risk of being hit from behind though.

The optimum filament drier just seems to be an air tight tupperware with 2cm of CaCl in a tray. Only the top 2cm absorb water. A fan might dry it out deeper. The mane trick is not penetrating the container with any wires, so the fan has to be powered by a Qi charger....Continue Reading

The LEDs were sliding off the sil pads & getting close to shorting out on the heat sinks so along came a new window design to try to hold the LEDs in 1 position. That seemed to work, but now the windows were big fat things hanging 11mm off the back.

The new windows are quite foggy. Solar filament drying hit a longevity snag. Making an air tight container which heats up, doesn't cost a fortune, & lasts more than 6 months is a super hard problem. It's going to be a long time before the windows are clear.

Another 1mm thicker to strap the switches in instead of hot snotting them. The trick with straps is making them slightly flatter so the welds press them down. Hot snot residue altered the switch dimensions. Everything else is going to need PLA welds instead of hot snotting. The latest thing is printing PLA carrier boards for the breakout boards. The carrier boards are CA glued on the container.

Now the rear panel is off 5mm because the switches were moved up 5mm.