state_tracker dropping knowledge too soon

[Wouter1]

Describe the bug
The state_tracker forgets too quick that there is some obstacle (a human agent) on the way
It re-routes the planned path, takes one step away from the human, and then turns around and steps again into the human.

To Reproduce
Run patrollingagent together with a humanbot.
Place humanbot at the right corridor somewhere in the narrow pathway.
Bots will run towards human , take 1 step backwards, and then one step forward, never getting around him.
Even with knowledge_decay=0 this remains this way

Expected behavior
remembering the obstacle, especially with knowledge_decay=0, and take a complete detour to reach destination

Additional context
I debugged this.
I think the problem can be localized to line 84 of state_tracker where objects are removed that
should be visible given the known perceptrange for that type of object.

This 'known perceptrange' info is stored in self.sense_capability and contains

{<class 'matrx.objects.agent_body.AgentBody'>: 2, <class 'builder.CollectableBlock'>: 2, '*': inf}

Because HumanAgentBrain is not in this list, it is assumed it should be visible anyway, and since it's not (it's just outside the range after the bot took 1 step back) it's removed.

Maybe the default is set wrong, or maybe this is a bug somewhere else. The sense_capability map is incomplete and the default is wrong in this case.

[thaije]

@Wouter1 thanks for notifying on this issue.

@jwaa could you take a look at this issue and #250, as the expert on the state tracker? 😄