Progress for week 39 (2019)

Fra Robin

(Forskjeller mellom versjoner)
Gå til: navigasjon, søk
(Wonho)
(Accounting)
Linje 8: Linje 8:
* Found out that 4D substrate is not supported in MultiNEAT
* Found out that 4D substrate is not supported in MultiNEAT
** Will need to modify source code. considering custom CPPN issue, this may be an option
** Will need to modify source code. considering custom CPPN issue, this may be an option
-
* Ran HyperNEAT with DyRET OpenAI environment
+
* Ran HyperNEAT with DyRET OpenAI environment [[https://robin.wiki.ifi.uio.no/Bruker:Wonhol#Experiment_result  gif]]
** simple 2D substrate, 4 input, 8 hidden, 12 output  
** simple 2D substrate, 4 input, 8 hidden, 12 output  
** crawling DyRET with something like hopping action after 20 gen
** crawling DyRET with something like hopping action after 20 gen

Versjonen fra 27. sep 2019 kl. 11:36

Innhold

Wonho

Budget

  • Find out how to customize CPPN in MultiNEAT - Risi has extra input parameter for leg length
  • Document setting up dev. env process
    • Look into Docker generation
  • Try simple HyperNEAT on DyRET env. - doesn't have to take account on leg length

Accounting

  • Found out that 4D substrate is not supported in MultiNEAT
    • Will need to modify source code. considering custom CPPN issue, this may be an option
  • Ran HyperNEAT with DyRET OpenAI environment [gif]
    • simple 2D substrate, 4 input, 8 hidden, 12 output
    • crawling DyRET with something like hopping action after 20 gen
    • Perhaps better result possible with assigning direction of joint rotation to output ( currently no adjustment done on either input or output to the network )

Tony

Budget

  • Get into a robot simulation program either webots, v-rep and etc...

Accounting

  • Have worked a lot on V-rep robot simulator. So far things have looked quite promising. Created my own 5-DOF manipulator --> This means I can either create my own unique robots or use other people's robots and test on it.
  • Managed to establish connection between server and client, i.e. V-rep allows one to send and receive commands between console windows. This is supported in Python which means I can take the outputs from neural network onto the robot with relative ease.
  • Have applied MDN-layer, a mixture density layer to the DNN to solve inverse kinematics which samples from probability distribution for example an x-input can have multiple values for y. Taking the average/mean will therefore be a problem, MDN solves this issue. Works perfectly fine in 2D, but with adding an extra dimension i.e. 'Z' and the neural model cannot estimate. Problem could be the loss function. Wll investigate this further next week.

Nikolai

Budget

  • Teste ulike tracking metoder
  • Lage et simpelt AR system


Accounting

  • Har laget en image tracking AR app til android.
  • Ikke klart å finne et robust tracking system som kan brukes i Unity.

Benjamin

Budget

  • Jobbe med datasettet
  • Lese på bruk av Bidirectional RNNs i Motion Capture.
Personlige verktøy