Empty goal returned after a fallen object is detected

[Hanziscool]

Pull commit 3080530..9a6b193 (minor clearing of comments) Bug:

1. object loaded and modelled with self.object_randomization=False (using same objects after reset).
2. objects falls
3. during the next Picking._reset_internal() cannot generate a new goal as this object has been removed from self.object_names (or possibly not_yet_considered_object_names).

Brainstorming:

• This bug probably is not dependent on whether we are doing object_randomization or not, simply that if we only have 1 object, and remove it, we return an empty goal...

[Hanziscool]

How to reproduce this bug :

1. put a brakepoint on L1250, when you reach there, overwrite the obj_pos[2] (as you did, but don't need to change the code, just do it directly in the 'output' window while you debug).
2. Then you should see how the relevant object is removed from self.object_names
3. Then further down, when you call get_goal_object() this will return an empty dictionary.
4. Then at the time we try to place it, there will be nothing there.

I am not quite sure if it matters if we have object_randomization.

• What we need to check (in both object_randomization = True or False ) is we have 1 object, if it is removed, will it get replaced to get a goal and be placed; or not.

[Hanziscool]

Current code:

• After detecting fallen obj, it will be removed from obj names, so now obj names is empty.
• then down there we will call get goal obj IF only self obj names is not empty

[Hanziscool]

That may not be enough.

1. I think self.object_placements is still not empty... and this list is used further down to place objects. Do we need to revisit how to update self.placement_initializer when an object is removed?
2. I would like you to just run the train_binPicking.py to see how these exceptions pop up...
3. I placed the train_binPicking.py:run_experiment_here.algorithm.train() around a try/except block with some code to let you debug from the terminal if there is a bug...

[Hanziscool]

after all objects fell, then it will be picked up by done and we will reset and we will always sample new object_placements in line 1146 before positioning objs in L 1151

[Hanziscool]

• The bug, as you see below, actually seems to be because object_placement is empty?
• object_placement may have an empty entry... where name is empty... is this coming from the get_goal_object return? If yes, we should probably not return anything in get_goal_object...
• need to follow through to make sure all logic is correct here.

[Hanziscool]

for relationalRL do:

1. get fetch
2. get pull origin picking_environment

This should help you get the latest changes.... as shown here:

• https://github.com/learningLogisticsLab/rlkit-relational/blob/picking_enviornment/examples/relationalrl/train_binPicking.py

[Hanziscool]

We can remove these empty assignments in get_goal_object and then test if that will cause any downstream problems, or perhaps more robust would be to check that the the name is not empty before trying to place anything....

[Hanziscool]

i think checking for not empty names is better to tackle other unforeseen cases where the obj.name might be empty

[Hanziscool]

I had put a bad check inside return_sorted_objs_to_model

Check for an empty goal

    if goal == {}:
        return {}

probably should be if goal == {} or goal['name']==[]

[Hanziscool]

check empty goal commit

[Hanziscool]

Still getting the same error... We are still missing something...

[Hanziscool]

• The reason is that we do not update placement initializers after an objects falls down.