search

cbfinn / gps

Guided Policy Search
http://rll.berkeley.edu/gps/
Other
597 stars 239 forks source link

Fixes for peg insertion BADMM task #23

Closed svlevine closed 8 years ago

svlevine commented 8 years ago

This PR contains a number of fixes. These fixes make the peg insertion task succeed (at least for me) on all four conditions. The fixes fall into a few categories:

  1. Changes to peg task to match the MATLAB reference:
  2. changed hole positions (still 0.2 apart, but center in a different place) -- no idea if this makes a difference
  3. changed cost function to raise the action cost and change wp
  4. Changes to algorithm parameters
  5. lowered dual_rate to 0.1 to match matlab reference
  6. changed max samples for GMM prior to match matlab reference
  7. raised # of SGD iterations (this seems to have made a difference)
  8. Changes to default algorithm settings
  9. lowered default polwt to 0.1 (1.0 is way too high for a default)
  10. changed the polwt lowering/raising heuristic to match matlab reference
  11. Deep copy bugfix
  12. added deep copy when flipping prev/cur variables for iteration, this appears to make a difference
cbfinn commented 8 years ago

other than the nit, LGTM!