Fritz Heckel » Weekly Update

Week 18 Status

heckel — Mon, 04 May 2009 14:47:58 +0000

Scientific:

Accomplishments:

Mark Overmars’ question about updates which increase the strength of influence points made a good point; the quick answer is that if you can put an upper limit on the expected strength of the point at insertion, we can maintain constant time updates, though sometimes regions will then have points with no strength in the local area.

Challenges:

There’s a question as to whether PRMs are a good choice for local navigation, so need to investigate some
For next influence points paper, or possible journal version, need to come up with 2 or 3 different concrete scenarios for using them, as some people were confused as to whether they would work in particular instances
Hierarchical influence points and splitting/joining of influence points needs to be dealt with.

Engineering:

Accomplishments:

AgentCore has some basics in, in that it will load up a list of agents from files and run them for a bit. There is no percept update, though effectors are generated.
AgentController is complete
Subsumption for ActionModel is to spec, mostly, though there may be minor issues with scoping
Loading agents from xml works
gotobehavior, alwaystrigger are complete
query file generation is mostly complete, though it doesn’t deal with ranges and enumerated options just yet
ThreadManager/WorkerManager works
A lot of behaviors/triggers added to Mantis
Created a model for BehaviorShop, and code to load an existing agent from XML
Wrote the code for BehaviorShop to load query files from XML

Problems:

Lots of behaviors and triggers to build
Need communication still, but giving a little more time before I implement my solution

Week 9 Status

heckel — Tue, 03 Mar 2009 15:04:30 +0000

Scientific:

Accomplishments:

Challenges:

Need to design!

Engineering:

Accomplishments:

Added approaching information to agent percepts, so it can now be used in triggers
Added speed of approach information to agent percepts; the accuracy is a bit bad, and I think this is related to the threading issues. Timestamps on the data coming out of FI3RST would probably also help.

Problems:

I can try to make everything episodic; I have a couple of ideas to try out. I’m guessing something simply time-based will work. PerceptManager can calculate these.
Agree with Dr. Y that we should probably go ahead and start using UDP for communication. Can try and add this to the python version now, or we can integrate it in 1.0. I’m concerned that we really just need to get working on 1.0 at this point.

I really think we should start work on 1.0, at least for BEHAVE. There are issues to address still, but I think we have to start work on the new version if we’re going to have it done or mostly done before GDC.

Week 7 Status

heckel — Mon, 16 Feb 2009 16:19:44 +0000

Scientific:

Accomplishments:

Challenges:

Engineering:

Accomplishments:

Integrated other people’s models
Integrated other people’s code
Cleaned up other people’s models
Cleaned up other people’s code

Problems:

Decomp still sucks, but at least we don’t walk off the world
Animations are there, they might be working. looped animations weren’t working yesterday, but Keith thinks he has them fixed now.
Finally started testing the astar pathing, but the gateway path code from Hunter is busted. Looks like the A* is correct. This is needed since the walkable building is a u-shaped obstacle
Colors in the marker system don’t work. Not even sure which version of the marker code to use
I’m sure there are other problems in BEHAVE that I haven’t gotten to test yet.

Week 6 Status

heckel — Wed, 11 Feb 2009 14:59:40 +0000

Scientific:

Accomplishments:

A lot of the basic ideas such as behavior parameters work

Challenges:

It’s hard to sense things like “suspicious” or “scared”. The temporary hack is to make these behavior parameters, and provide agents with limited ability to sense a few behavior paramaters, but this needs some thought
Sometimes you just need inventory and other basic game-ish things. We’re going to have to add in some basic features to the agents
Desperately need to start using additional tabs in BehaviorShop. We’re putting stuff into the main tab which belongs under dialog, and other such things. I also suspect that we need to add some study infrastructure for behavior shop.

Engineering:

Accomplishments:

Multiple triggers definitely work
The behavior parameters definitely work.
Agent parameters, reflecting “opinions” and relationships with other agents, appear to be working
We have some decent reference agents, though they’re not all complete

Problems:

Need animation integration
Need fixed up decomps
C++ rewrite still on hold

Papers

FDG and FLAIRS camera ready copies coming up soon
Sandbox paper has been outlined

Week 4 Status

heckel — Wed, 28 Jan 2009 14:27:37 +0000

Scientific:

Accomplishments:

A working plan for dealing with the behaviors needed for the guard scenario was sorted out

Challenges:

During the rewrite, need to go over the engine pretty carefully and make sure we’re dealing with memory, agent relationships, and other such things in a principled way. We’ll need to start using the other tabs in BehaviorShop

Engineering:

Accomplishments:

The difficult behaviors are pretty much in place and ready for testing

Problems:

Haven’t really touched the C++ rewrite this week
At this point, need the animations and a good way to actually invoke them; since that part of FI3RST has been a bit of a mess, getting them working has been painful. I’d really rather not be writing the timing logic to make sure they don’t get interrupted into the behaviors.
Currently have written some things into behaviors which belong in other sections of the agent description

Week 3 Status

heckel — Fri, 23 Jan 2009 18:58:42 +0000

An admittedly unproductive week, and I don’t have a good excuse.

Scientific:

Accomplishments:

FLAIRS paper accepted

Challenges:

Not happy with how I’m storing information about other agents yet

Engineering:

Accomplishments:

Agent parameters, for storing knowledge about other agents, are hacked into the perceptual model
Behavior parameters hacked in
Triggers on parameters are mostly in place

Problems:

Still working on SSPS rewrite
Working on new behaviors for guard scenario
My original percept passing method is really too ugly to live, but I’m leaving it in place since we’re migrating to C++ anyway

Papers

IROS paper possibility for March 1

Week 1/2 Status

heckel — Mon, 12 Jan 2009 19:17:16 +0000

Scientific:

Accomplishments:

FDG 09 paper got out on the 23rd
HRI 09 abstract has been accepted. I’m author 4 of 6 on this paper/poster, for setting up the experiment while at NRL and drafting the first paper

Challenges:

Influence points need to be integrated into BEHAVEngine properly
Behavior parameters should probably be formalized; they’re simple, but I want to be sure I’m using them in a principled manner

Engineering:

Accomplishments:

C++ rewrite of BEHAVEngine and associated utilities begun; it should end up being fully cross-platform
Persuaded all necessary libraries to work : boost, stlport, libxml
agentcore and agent XML files fully parsed in C++
problems in agent XML schema fixed

Problems:

Need to figure out how to use boost graph library, which should make working with SSPS easier and more efficient
SSPS rewrite shouldn’t take long, but will take a bit of time
Hunter reports some sort of crash in Python BEHAVEngine, but I have no details at present
Generally just need to do the BEHAVEngine rewrite, including Influence Points and all behaviors

Papers

IAAI paper on DASSIEs (Dr. Y as lead)
IROS paper possibility for March 1
GDC scholarship deadline is soon

Week 51 Status

heckel — Wed, 17 Dec 2008 16:42:32 +0000

Scientific:

Accomplishments:

My brain still works

Challenges:

Need to fully flesh out influence points for the FDG paper

Engineering:

Accomplishments:

Reworked the percepts so that affordances, percepts, transients, etc. are all consistent. Percept objects are now always dictionaries with a particular set of keys.
Built the dynamic object registry, which keeps track of all current agent locations as well as object locations
Added percept types: object (objects and agents), comm (communication strings between agents)
Rewrote SSPS. I feel like there’s no need for SSPS to be a separate service running in its own thread. There are some optimizations to be made though
Added capability to add and remove objects from the world to BEHAVE; this relies on having a copy of the World object
Helped Hunter draw DEACCON regions in the world for debugging
Added flag drawing on region gateways to the world
Wrote up baseline bomber and bomb squad xml behaviors
new triggers: find object, time
new behaviors: explore, add object, remove object, guard area
behavior respiffs: goto, patrol, wander (to use SSPS)
action model now holds onto copy of SSPS object so it can navigate the world– behaviors only need to give a goal point, action model finds the way there.
changes to trigger so that they all have an “active” instance variable and a deactivate method; not all triggers have been changed over to this way
my D* implementation works now
Probably more. I’ll update if i remember things I forgot later.

Problems:

SSPS could use some optimizations, probably hinting for find region calls– give the last known region, and ssps can first check neighboring regions of that region, and then do exhaustive
SSPS has a somewhat dodgy A* implementation which will need to be adjusted, but this can wait until influence points and behavior parameters are in. once we have a more dynamic world, it makes sense to switch this to a D* variant, but it’s not necessary now
Behavior parameters are sort of half in, but I haven’t decided the most principled way to apply them still
DOR stores objects and agents separately, which is fine, but stores each type in simple lists. this is okay for now, but ultimately needs to use the region decomps
animations are dicey– needs a rework in FI3RST. possibly ignore further commands while special animations are running
the consistent dictionaries for percepts are an improvement, and give me some flexibility right now that i need, but percepts should be objects. i think combining all the information types into a unified percept type is the right thing though
the camera in FI3RST badly needs an overhaul. no idea where to set the culling plane. camera should also start out unattached in the world so that agentcore requests all the NPC agents that are created. i have no control over the model of the first agent because of this
we’ve got to be a bit more organized about our world scale and creation. if deaccon can’t handle the current size, it should really be switched over to being based on doubles rather than floats, or it should scale the world up, operate, then scale everything back down. this was a real mess
the BEHAVEngine is a bit of a mess right now– data doesn’t move through it in a clean way, which sometimes makes adding features difficult. of course, most of the time i don’t expect to be adding major new features, so some brainstrain is to be expected when making modifications on the scale of all the above
I need to know more about the pipeline; it was hard to know who I needed to do what at times this past weekend

Papers

FDG paper. Title uncertain at this moment, but will figure something out soon. Clearly on influence points.
Doctoral consortium submission.. I can’t say I’ve even started thinking about this, though I did get briefly excited after either hallucinating or reading a typo the post by Dr. Y’s door that said the 29th was the DC deadline

Week 47 Status

heckel — Wed, 19 Nov 2008 14:16:39 +0000

Scientific:

Accomplishments:

Not very much, though the work with the affordances at AAAI FSS gives a good basis for influence points, since they can work in a very similar manner to affordances.

Challenges:

May need to get a Hofstede book through ILL; they’re all checked out for a month or more at the library right now.

Engineering:

Accomplishments:

Cleaned up the OpenSteer code so that we’re just using FlexiVehicle, next I’ll need to make a sim/factory class for it so that we keep track of all obstacles, agents in one place for the sake of proper obstacle avoidance.
Also all the work from AAAI which I did not write up last week– affordances, agent scaling..

Problems:

Still don’t have world geometry
Turns out that I was misled by the internet. BEHAVE will need to go to Python 2.6 and use the multiprocessing module, which I expect to be a moderate pain in the rear.
My D* implementation is stubbornly refusing to work properly

Papers

FLAIRS paper w/ user study isn’t feasible without world geometry from SSPS and world. I need to have this at least a week before we start a study to be certain everything works
Alternate FLAIRs paper possibilities: Description of the architecture– not really ideal, but it would work, and we have everything we need for that. Influence Points – without world geometry, we can’t really present more than talk on this. At least I know now that they can be treated in pretty much the same way as affordances, based on how the affordances are set up now.

Week 44 Status

heckel — Tue, 28 Oct 2008 17:46:48 +0000

Scientific:

Accomplishments:

As part of the Intro to Research class, we had to write up a problem statement and a list of papers for a literature review. It’s a little short yet, but I’ve written it up here, and I’ll probably just keep that page around as a literature list. Also, the question isn’t quite right yet.

Challenges:

While I have a temporary solution for this cycle of development, there is still a major question of how to deal with percepts properly. I really want to talk to a few people at the fall symposium about this, since I’d like to aim for a solution that is somewhere between an efficient architecture approach and a more realistic cognitive approach.
The hearnoise trigger has a hack, as I believe I mentioned last week, which keeps it active until the behavior it triggers has completed its task. This is clearly not the right approach, so this is a place where we’ll really want a hierarchical behavior that uses some internal state to set goals. This is all tied up with the behavior parameters, I think, at least insofar as it’s all agent state.

Engineering:

Accomplishments:

To help resolve the percept issues temporarily and permanently, I’ve introduced the PerceptManager. Each agent controller thread gets a perceptmanager that passes on the relevant percepts (so it will filter based on whether the percept is visible), and maintains a bin for pure percepts (straight overwrite) and a queue for transient percepts. It still needs a little work, but is basically set.

Problems:

See last week.