Inqster - The Next Generation

Inqster in the Flesh

Finished the mechanical design and printings.  There might be some miscellaneous parts to attach electronics, but this is the relatively fixed mechanical chassis.

...and here is the printed chassis.

Here is a current weight sheet.  I've added plastic and component cost of the expected completed Inqster.

Inqster will be a test mule.  No beauty contests will be entered.

@inq Where do you get an ESP32-S3 for $4.96? Amazon is in the $30 range, eBay is $7.74. 

I just ordered 2 esp32-s3 and 2 esp32-c3 with shipping for $47.41 CDN / $35.28 USD on AliExpress.

Posted by: @zander

↑

@inq Where do you get an ESP32-S3 for $4.96? Amazon is in the $30 range, eBay is $7.74. 

I just ordered 2 esp32-s3 and 2 esp32-c3 with shipping for $47.41 CDN / $35.28 USD on AliExpress.

https://www.ebay.com/itm/175731198587

Sorry, I miss-spoke.  I forgot I got it on ebay and it had shipping extra.    They're $6.94 with shipping.

@inq
Inqster will be a test mule. No beauty contests will be entered.

The colors you used reminded me of all the pastel colors I saw being used in the Barbie movie promos so I guess it will win a Beauty contest 🙂

One issue I found outdoors was going over an obstacle. If the obstacle was 1/3 the diameter of the powered wheels the robot it could easily pull itself over although it was wise to have it align both wheels to the obstacle to prevent lying over at an angle. If the swivel wheel was going first it simply swiveled and jammed on just about any solid edge over or even under the 1/3 swivel wheel diameter. The robot base had no trouble pulling itself through softish ground or grass if the powered wheels were in front not so much for the swivel wheel in front. My base was many times heavier with a steel frame holding heavy motors and battery but your base looks very light so maybe yours will be different it will probably bounce along like a kangaroo when it hits a bump 🙂

Posted by: @robotbuilder

↑

@inq
Inqster will be a test mule. No beauty contests will be entered.

 

The colors you used reminded me of all the pastel colors I saw being used in the Barbie movie promos so I guess it will win a Beauty contest 🙂

One issue I found outdoors was going over an obstacle. If the obstacle was 1/3 the diameter of the powered wheels the robot it could easily pull itself over although it was wise to have it align both wheels to the obstacle to prevent lying over at an angle. If the swivel wheel was going first it simply swiveled and jammed on just about any solid edge over or even under the 1/3 swivel wheel diameter. The robot base had no trouble pulling itself through softish ground or grass if the powered wheels were in front not so much for the swivel wheel in front. My base was many times heavier with a steel frame holding heavy motors and battery but your base looks very light so maybe yours will be different it will probably bounce along like a kangaroo when it hits a bump 🙂

Colors - Interesting how all those 18650 cells are out of Dell laptop bricks.  They must use vendors from all over the place.  Most are good Panasonic cells, but other aren't.

I definitely plan on it being a tail-dragger.  I'm still project the all up weight being less than 2 Kg.  Right now with everything except electronics, it's about 1800 grams.  I think it will do pretty good around the yard and gravel driveway... at least at walking speeds.  We'll have to see what it's threshold of pain is. 😉 The main suspension has about 4 cm of travel between off the ground and sitting on the tires (as in the picture).  The suspension can be moved fore/aft on the battery bank on holes like the old erector sets, if I need to load/un-load the tailwheel.  I'm shooting for an old Cadillac ride over a kangaroo.

@inq Still cheaper than what I found.

Posted by: @zander

↑

@inq Still cheaper than what I found.

I noted that the specifications are all over the place for flash size and PSRam size.  Definitely not like the ESP8266 boards.  I'd imagine the prices would be accordingly different.  For instance the ones I posted have 16MB of flash and 8MB of PSRam.  Many of the cheaper ones, seem to be doing 8MB/2MB.  You have to really inspect them to see what you're getting. 

@inq I have to admit I didn't pay attention to the details, and I should. The C3 is N16R8 but the S3 is only N8R2. I have messaged them to substitute or cancel the S3 and I have a pair of s3 N16R8 in my cart ready to go AND a few pennies less. I am definitely slowing down in my old age.

Posted by: @zander

↑

@inq I have to admit I didn't pay attention to the details, and I should. The C3 is N16R8 but the S3 is only N8R2. I have messaged them to substitute or cancel the S3 and I have a pair of s3 N16R8 in my cart ready to go AND a few pennies less. I am definitely slowing down in my old age.

Well, as you know... not all projects need lots of flash, or even ram.  Heck! as I've contended I've been able to do some pretty sophisticated projects on a mere ESP8266.  And I'm still a little fuzzy on what I'm going to do with the PSRam.  😆 I just accidentally found that chart when wondering what all the numbers at the end meant.  I wanted the larger one and I wasn't able to find a 32MB version.  I definitely use the flash as a disk for web content. 

Although not really knowing for sure what I'm talking about (yet) it appears to me the Neural Network concept learns a matrix of weights and biases.  The data input (sensors) use linear algebra to run through the matrices and finally gives values to be used on the drive motors. 

After learning, instead of hard-coding those weights and biases, it would seem best to also store them on the disk.  I also believe as the goal changes, the learned weights and biases would also be different.  For instance the goal to maximize speed, minimize turning and maximize distance from obstructions that we were talking about earlier in the thread would be completely different from the learned values when wanting to move close to ALL the walls and obstructions so that we can map the building.  Those weights and biases would be completely different.  Therefore, I'm thinking I'll need several sets of learned matrices depending on the current task at hand.  

In addition, I want to map the building.  I imagine all that data may have to be sent to a base station via WiFi.  But I might want buffer it and send in blocks.  Either way, that too will take plenty of disk space.  I wanted the largest one I could find.  I guess I could always add an SD card locally.  Anyway... plenty of options as I learn more and things play out.

@inq If you allow the observations of the village idiot when it comes to all this artificial intelligence stuff, I couldn't help but notice that when I was watching a YT video about ChatGPT, they used all the same terms you and others used when discussing your robot. Is there a way to use ChatGPT in the robot 'brain'?

@zander wrote:

 "Is there a way to use ChatGPT in the robot 'brain'?"

The boffins are working on it.

https://singularityhub.com/2023/08/02/deepminds-chatgpt-like-brain-for-robots-lets-them-learn-from-the-internet/

https://www.deepmind.com/blog/rt-2-new-model-translates-vision-and-language-into-action

@robotbuilder The problem is that first link says 'internet' ChatGPT is NOT connected to the internet, it's cousin WolfRam Alpha is.

@zander 

I read that is about to change.

If one day they have a simple plug in brain for robots my interest in robot control systems will cease not that I will be around to see it.

Posted by: @inq

↑

After learning, instead of hard-coding those weights and biases, it would seem best to also store them on the disk.  I also believe as the goal changes, the learned weights and biases would also be different.  For instance the goal to maximize speed, minimize turning and maximize distance from obstructions that we were talking about earlier in the thread would be completely different from the learned values when wanting to move close to ALL the walls and obstructions so that we can map the building.  Those weights and biases would be completely different.  Therefore, I'm thinking I'll need several sets of learned matrices depending on the current task at hand.  

In addition, I want to map the building.  I imagine all that data may have to be sent to a base station via WiFi.  But I might want buffer it and send in blocks.  Either way, that too will take plenty of disk space.  I wanted the largest one I could find.  I guess I could always add an SD card locally.  Anyway... plenty of options as I learn more and things play out.

Have you considered a bank of EEPROMs ? Rather than using disk(s) or SD card(s), you may be able to store the data locally (and persistently) on a few EEPROMs. Say one or two that contain your different sets of weights for navigating quickly and bump-free inside the house, another for mapping (maybe just the left-hand-rule for maze solving) and another couple for outdoor use (one for mapping the lawn and another for avoiding goats).

You could save the maps generated for your house, the gym, the yard and so on on them as well. You'd have access to all your data and all your characteristics and could easily switch between locations and behaviours by reloading your map and your weight set. I believe EEPROM should be safe since you'll only be writing to it a few times and reading it forever afterwards (so it won't wear out), it's faster than SD cards and requires less power than a disk.

@will 

My view is that it will turn out harder than you think using networks then simply hard coding these behaviors. The genetic algorithm example was just that, an example, a tutorial, using a restricted situation to demonstrate how to code the weights in genes and trial hundreds of robots and then selecting the genes of the best performers to generate another set of weights.