#robotics Logs

Aug 21 2014

00:44 MrC123 oooh ka ooh ka ooh ka shaka
00:44 rue_bed use about 10%
00:45 rue_bed meaning to mimmic the human brain you only need
00:45 GargantuaSauce the 10% thing is total tripe
00:45 rue_bed 9 billion nurons worth
00:45 rue_bed 9G
00:46 MrC123 so taking the 10% we use, losing the 90% we dont, what are you saying, a cat or ferret sized head is all that is needed
00:46 MrC123 ferret...in..... space!!!!!!!
00:46 rue_bed no, I'm using it to justify how much computing space is needed for simulation
00:47 MrC123 simulate in earnest.... just allow it to run at a slower iteration rate than live...
00:47 rue_bed heard of the fellow who had a brian tumor, shot himself in the head and only blew out the tumor, lived
00:48 MrC123 there was the guy with a railroad spike in his head... another with a steel rod through his head... lots of cases of brain penetration that were not fatal
00:49 rue_bed see? 10%
00:49 rue_bed maybe the other 80% is redendancy?
00:49 rue_bed :)
00:50 MrC123 i dont think it is 10 % start to finish... i think it is 100% available pathways, refined down to 10% over course of learning and life
00:50 MrC123 i.e. its a multi level neural network, that after training can be optimized via 90% neuron removal
00:51 MrC123 but until the training is laid in, the entire 100% is critical as it is impossible to predict the pathways selected through learning and optimization....
00:52 rue_bed hmm
00:53 GargantuaSauce that dude also underwent a severe personality change as a result of that incident
00:53 MrC123 now if you could take a mature intilect, you could read, optimize, and replicate using only 10% totaul neurons
00:53 GargantuaSauce anyway you arent going to make a successful foray into neuroscience with the axioms you've mentioned
00:54 MrC123 rue: can you pm or repeat the axioms so i can join in (this time in sync)
00:55 GargantuaSauce you didnt miss anything
00:56 MrC123 ty
00:56 GargantuaSauce but seriously
00:56 GargantuaSauce the whole 10% thing
00:56 GargantuaSauce is 100% bullshit
00:57 MrC123 now i suppose it might be interesting to start with a small neuron set, then introduce a rolling set of neurons that have a live span to get trained, then weed out the useless, grab the good into the main, repeat...
00:58 MrC123 so every interval, 1k neutrons in a N-level net are introduced, trained, and then optimized, then combined with base, ,and repeat
00:59 MrC123 someone call BS and ask me to STFU please!
00:59 GargantuaSauce that's not how a nn works
00:59 GargantuaSauce stfu
01:00 MrC123 by interval i mean period of time to tweak the neuron values to hone in on the local minima, not iteration as inone crank at the trainhandle...
01:01 MrC123 so one interval would = 10k iterations in a backward propagating neural network of N layers with positive reinforcement
01:01 MrC123 NOW i will stfu
01:05 C123 hands the mikes to rue and gar and tries to turn down the feed
01:50 MrC123 so while neural networks and rule based systems are good for solving problems....
01:50 MrC123 how do you create a construct that seeks out the next problem on its own
14:18 deshipu I wonder what would make more sense, when asked, send the current status of the accelerometer, or the average since the last query
14:20 deshipu or maybe the average of last n measurements
14:44 Tom_itx average
15:09 deshipu Tom_itx: since last measurement or from the last n milliseconds?
15:11 Tom_itx depends what you're looking for i suppose
15:12 Tom_itx if you want a current reading, i'd take a few samples and average
15:12 Tom_itx i suppose the numbers bounce around quite a bit
15:13 fluffywolf depends on what you're doing. navigation? average. detecting you just hit something? instantaneous.
15:13 fluffywolf and, yes, your numbers might be very noisy.
15:15 fluffywolf no idea
15:15 fluffywolf grr, wrong window
15:18 deshipu well, I have several uses in mind, I'm mostly afraid that I will miss an event between readings
15:19 deshipu taking averages over longer time would alleviate that
15:19 deshipu so, one thing I want to do, is to detect when the robot is picked up from the floor, and start to wiggle all the legs randomly then
15:19 fluffywolf you can also store both the average since the last reading and the maximum since the last reading
15:19 deshipu and then detect when it's being put back (or dropped) and spread the legs for safe langing then
15:20 fluffywolf you should have hardware signal processing (i.e. a RC lowpass filter) sufficient to average between a/d samples.
15:20 deshipu second thing is detection of collisions with objecst while walking -- but only, when moving forward one of the front legs
15:21 deshipu and finally changing the walking direction based on the slope -- walking uphill or downhill
15:22 fluffywolf store min, avg, and max, if you're worried about missing something. being picked up will be an increase in the vertical, done picking up will decrease then return to g, dropped with be 0, etc.
15:24 fluffywolf RustyShackleford: time to build your autodialer, then. :P
15:25 fluffywolf grrr. stupid irc client
15:25 deshipu fluffywolf: great idea, thanks
15:26 fluffywolf hitting something will be a high peak horizontal...
15:27 deshipu sounds like I will need to implement some kind of pattern matching :)
15:29 fluffywolf yes. :P
15:31 fluffywolf you could do fancy AI, or you could just have a block of if statements (if(accv > g*1.2) pickedup=1; if(accv < g*0.8) donepickedup=1; if(accv < g*0.1) dropped();...).
15:32 fluffywolf I'd suggest a block of if statements for your first version. :)
15:32 fluffywolf can your legs actually absorb the impact of being dropped in a non-permanantly-deforming fashion?
15:32 deshipu a state machine
15:33 deshipu yes, they are pretty flexible, they are basically just 3 servo horns screwed together
15:33 fluffywolf yeah. that'll work about once...
15:33 deshipu it's very light
15:34 fluffywolf does it have some logic to always land on its feet? :)
15:34 deshipu basically just the 12 9g servos an a small lipo
15:35 deshipu that would be extremely cool, but for now I just want to spread them while falling so that it falls roughly on them
15:35 deshipu I might record a 'flipping the right way up' routine later
15:35 fluffywolf note that simply detecting which way is down when in freefall is non-trivial.
15:35 deshipu there is so much fun stuff that can be done with a quadruped and an accelerometer
15:36 deshipu fluffywolf: it's not non-trivial, it's impossible with accelerometer
15:37 deshipu well, there is some air drag...
15:37 deshipu but that's too small
15:37 fluffywolf it's quite possible... you just have to integrate your orientation from the last known position... inertial navigation.
15:37 fluffywolf but it's not trivial to make that work.
15:37 deshipu ugh
15:37 deshipu I really don't want to go that way
16:20 Hyratel http://www.mini-box.com/OpenUPS2
22:30 rue_shop2 tried another casting, bigtime fail
22:32 rue_shop2 hmm not what I'd call a ups...
22:32 rue_shop2 why tho
22:32 rue_shop2 its just a battery backup with switching regulator
22:32 rue_shop2 its the sum value of $5 on ebay
22:33 rue_shop2 hahah they cant even supply the batteries, thats $3 of the $5 value
22:35 rue_shop2 no sign of the small cat
22:35 rue_shop2 :(