Thanks Alexander. My Pong project was similar in certain aspects. The
first thing that i wanted to do was predict where the puck/ball would be
when the paddle could make contact. Instead of NuPIC i used an R*Tree
that remembered (literally) the paths of the ball that it had seen
before. I planned on generalizing this, but it worked fine for testing
at the time.
The problem that i had was in moving the paddle to the predicted
intersection point (as i described in the blog). It looks like TuxPuck
moves the bat at a constant velocity. I was trying to create movement
plans that were more biological.
On 7/12/2014 1:28 AM, Alexander Hirner wrote:
On the notion of preconceiving the position of your pong paddle, the
work of Steve Levis might come in handy. He did a fully integrated
sensory-motor loop with TuxPuck at the Spring Hackathon. However,
NuPIC is only used to augment the predicted location of impact of the
puck afair.
http://numenta.org/blog/#NuPUCK
On 12.07.2014, at 01:42, Matthew Lohbihler <[email protected]
<mailto:[email protected]>> wrote:
Hi Ian,
I have a couple experiments. Both use JBox2D as a simulation
environment. The first one i called "Pong+" because it's the classic
game but with ball spin, air friction, and round "paddles". (The
enhancements are meant to prevent agents from trying to reverse
engineer the environment, but rather make necessary a solution based
upon a feedback loop.) A relatively long-winded description of the
problems that i encountered is here:
http://blog.serotoninsoftware.com/steps-toward-a-solution, and here:
http://blog.serotoninsoftware.com/box2d-machines.
Due to those problems i started the second experiment, which is a
model of a leg. I'm trying to get it to balance, bend, and hopefully
in time, jump and walk (with the help of a second leg). The problem
that i'm having this time is that inputs are position and velocity,
while the actuators are forces, which are at a different differential
level. I'm hoping to solve this by using time steps to convert the
velocities into accelerations, but i haven't done any testing yet.
Paying work always rudely gets in the way.
Cheers,
m@
On 7/11/2014 12:11 PM, Ian Danforth wrote:
Matt,
I've used physical robots for this, but haven't gotten very far
using just NuPIC. It's not set up to handle the sensory-motor loop
at all. I'd love to hear more about your setup!
Ian
On Fri, Jul 11, 2014 at 8:54 AM, Matthew Lohbihler
<[email protected] <mailto:[email protected]>> wrote:
Did i maybe send this to the wrong mailing list? Or is the
question premature?
On 7/8/2014 3:50 PM, Matthew Lohbihler wrote:
Hi everyone,
I finally watched Jeff's presentation on sensory-motor
integration the other day, and thought it was great. After
thinking about it a little bit though, it occurred to me that,
to implement, you'll need sensory data (which can come from an
environmental simulation), but also motor data (i.e. the innate
behaviour of the agent). I was curious about what test
environments you are using and how you are generating motor data.
For my part i've been using JBox2D to create simulated
environments, and have been working on building innate motor
skills that look biologically realistic, but it's not so easy
to do.
Regards,
Matthew Lohbihler
_______________________________________________
nupic mailing list
[email protected] <mailto:[email protected]>
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________
nupic mailing list
[email protected] <mailto:[email protected]>
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
