Information related to the GIVE 2.5 efforts for the 2011 submission at the Grupo PLN FaMAF.

People

  • Luciana
  • DrDub
  • David

Topic: Grounding

Tag and branch creation in the repo

To create a tag with the contents of trunk's revision X do the following:

svn copy https://csystem.googlecode.com/svn/trunk https://csystem.googlecode/svn/tags/vX

To create a branch named 'new_branch' do:

svn copy https://csystem.googlecode.com/svn/trunk https://csystem.googlecode.com/svn/branches/new_branch

Ideas

Put your name before an idea if you support it. Put it (also) after if you want to help implement it.

  • Extended plan descriptions
    • David DrDub Focus on providing a very good 'H' (help) command. DrDub David
    • David DrDub Focus on introductory descriptions of the overall goals, map and actions, stressing dangers and challenges. DrDub
    • David DrDub Ground safes telling the player that the trophy is inside (or that not is) and describing how many buttons left to open it. David
  • Techniques
    • DrDub Use reinforcement learning
    • DrDub EM using human instructions DrDub
    • David DrDub Do aggregation at the CNLG level (is there a need?)
      • Maximum number of messages is 3.
      • There's a need, but it is in tension with having short messages.
    • DrDub Wizard of Oz using speech to gather some ideas DrDub
  • Infrastructure
    • DrDub Implement a tool for designing new worlds.
    • David DrDub Deployment of GIVE tagged systems + tagged log acquisition.
  • Grounding
    • David DrDub Ground alarms as soon as one becomes visible (using distance too). David
    • David Luciana In the deduction by elimination strategy use yes and no. Say "Yes! Push that" instead of "This one" Luciana
  • Testing
    • Design a map with a polygon with many booths that are a single room internally.
    • David DrDub Make sure the "matrix RE" algorithm doesn't fire in non-matrix situations. David
    • David DrDub Automated test cases for GIVE (random maps, random states, check for exception and spurious outputs) DrDub
  • General improvements
    • Consider the alarms as walls?
    • David DrDub Use questionaire data for verbalizations DrDub
    • David DrDub Continuous replanning in a separate thread DrDub
    • David Use landmarks for RE. David

Issues in Current System

Add the issue, include your name and date (you can do that by writing 4 tildes '~')

  • BUGS:
    • The system gets mute after "thinking" that the last instruction sent was not read Luciana 18:58, 28 April 2011 (UTC)

Sent:{Press the red button in front of you., 3500, 0, manipulate(b13,off,on,reg_b13_b14_room-1)} HANDLE_STATUS_INFO ---- Lost:{Press the red button in front of you., 3500, 0, manipulate(b13,off,on,reg_b13_b14_room-1)} HANDLE_STATUS_INFO ----

    • Since rev 18 the system raises an event of type TIMEOUT every 7 seconds. Besides, it considers that the last instruction understood by the player is the last generated and sent instruction which has a related plan step. David 22:26, 4 May 2011 (UTC)
  • Any problems with "move to the center of region" strategy? DrDub 17:07, 22 April 2011 (UTC)
    • There are few cases were this is an issue (diagram to follow), but it is annoying. The main problem is when the object of interest is visible then it should change from going to the center of the room to the button region. Possible solution: don't use the center of region but the point of the region closer to the next region in the plan and visible. Luciana 17:32, 28 April 2011 (UTC)
    • When the object of interest is visible the object can be directly referred. In other words, the idea is to implement a variable radio depending on the number of visible distractors. Luciana 17:32, 28 April 2011 (UTC)
    • When the object of interest is visible but is not too close to the player the system tells the player to go to the button's region center that is a small region that wraps the button. David 00:18, 24 April 2011 (UTC)
    • This could be resolved by saying to the player to move in direction to the point that's visible and that's closer to the next region. Not done yet. David 14:39, 13 May 2011 (UTC)
  • Change 'one of these' to 'one of these, get closer and I'll tell you which exact one' DrDub 17:07, 22 April 2011 (UTC)
    • Done. I change the 'one of these' instruction to 'One of these. Get closer' because is shorter than the message that Pablo suggested. David 14:39, 13 May 2011 (UTC)
  • The system is not reactive, if the user does everything in less than 200ms, the user will be idling her thumbs. DrDub 18:01, 22 April 2011 (UTC)
    • A solution for this is to ensure that the granularity is more than 200ms.
    • Another solution is to make this timing adaptive.
    • Watch-out, the 200ms comes from the GIVE framework, so to change this you'll need a separate thread to compute things from the world.
    • Frequently the system does not say anything for periods much longer than 200ms (for more than 2 seconds, for sure). Why is this? Luciana 20:36, 23 April 2011 (UTC)
    • The system could mute but not for more than 7 seg because an TIMEOUT event is launched. TIMEOUT events are not launched when the user is in "deduction by elimination" phase, so the system could mute in this case if the player stay without move or turn in such a manner that the list of visible objects maintain constant. David 00:18, 24 April 2011 (UTC)
  • The system went mute after I didn't do anything for a few minutes, pressing H didn't work either Luciana 20:48, 23 April 2011 (UTC)
    • In the branch I implemented two types of timeout's events. One called MOVE_TIMEOUT that Monitor raises when the player hasn't done the last plan step that the system verbalized in 7 seconds. The other is called MANIPULATE_TIMEOUT and it's raised when the player hasn't manipulate the target object in 7 seconds. I know this is not the root of the problem but I fixed another things and probably this bug is fixed too but I am not sure. David 18:33, 3 May 2011 (UTC)
  • I started doing actions before the system finished with the canned text. It seems that the system did not replan in order to consider the current context. It gave me the first instruction of its original plan which was certainly outdated, and then I think that the system broke because it went mute. Luciana 20:54, 23 April 2011 (UTC)
    • Fixed in the branch. The system shows the welcome messages and, at the same time, gives a feedback if the player press a button or sees an alarm. Then continues showing the remaining welcome messages. After it finishes with the welcome, it replans and start with the instructions.
  • La version actual de repo (r8) no funciona con los ultimos tools de give (v2.3.3). El error es que se cae el sistema NLG diciendo "connection reset". Luciana 22:36, 23 April 2011 (UTC)
    • Fixed in version 9 (r9). We had to change the nlgsystem dependence in the pom.xml file. David 20:52, 24 April 2011 (UTC)
  • There are urgent messages, alarm warning and wrong button pressing, that are being replaced by not urgent messages. Implement the distinction between not urgent messages and urgent. Urgent messages cannot be replaced by not-urgent messages.Luciana 18:21, 28 April 2011 (UTC). Do not replace push instructions.
    • Fixed in the branch. I added a priority to each message. If the system is showing msgs A with priority p1 and the msgs B with priority p2 is generated, then A is replaced for B if p1<=p2. The messages can be showed in a 'block'. When two messages A and B with priority p1 and p2 respec p1<=p2. are added in this way, then the system shows A before B without any replacement. DavidRacca 18:15, 3 May 2011 (UTC)

Meetings

18/03/2011

Present: Luciana, DrDub,David

Discussed generalities regarding David's thesis, bounced the idea of doing a "blind" GIVE for NLG Instructions for the Visually Impaired. Decided to have weekly meetings on Thursdays at 3PM.

31/03/2011

Present: Luciana, DrDub,David

Discussed some ideas for the challenge. These are in the ideas' list.

07/04/2011

Present: Luciana, DrDub,David, and Alexandre Denis

  • Alex presents the NA system.
  • We play with the NA system.
  • Suggestions and modifications to the agenda more than welcome :-).
  • Reminder: give Clark's book to David & Pablo on a pendrive.

19/04/2011

Present: Luciana, DrDub,David

  • Determination of GIVE deadlines.
  • Thought about how to package the java classes of the project.

22/04/2011

Present: DrDub, David

  • Pablo configured and created tag version 0 to run in cheron server.
  • David thought how to solve the alarm problem. This is when the system takes the player over a way in which there is an alarm. David realized that the actual system's strategy should work well using the overAlarm method of CWorld's class except the case in which the player is in a room that has several alarms. This exception is a consequence of "center of region problem".