blocks coordinates (This is actually about the NLP task)

[MrRobot2211]

Hi. where could I find the exact definitions of the world parameters for the NLP task

Particularly what is the meaning of this array gridworld['worldEndingState']['blocks'].
Do the coordinates represent position and colour of each block?
If so how. Which are the limits? which is the colour mapping? How is it related to the position and the direction of observation or are they absolute?

[artemZholus]

Hi @MrRobot2211 , yes, the gridworld['worldEndingState']['blocks'] is a list of 4-tuples representing 3-d coordinates of a block and its color. For technical details, I recommend you to look at the gridworld repo: https://github.com/iglu-contest/gridworld
It provides the training API for the RL task, however, it is also useful for the NLP task. In particular, there are parsers for each dataset we have. E.g. https://github.com/iglu-contest/gridworld/blob/master/examples/iglu_dataset.py

Please let me know if this answers your question.

[MrRobot2211]

Hi Artem thank you I will take a look.

Regarding the blocks specifically I am confused . Which coordinate is the colour? And which is the coordinate order?

I am confused because I checked and there are two coordinates that achieve negative values, the ones in positions 0 and 2 . And 1 and 3 look like position coordinates because they have values over 60, and per the documentation there are only 7 colours from 0 to 6.

[MrRobot2211]

For clarification, I am attaching the statistics over all the block values for each coordinate

[MrRobot2211]

Another point I am confused about here is that the size of the grid is per the documentation (9, 11, 11 ) but then how can the coordinate position be over 60 is there a rescaling somewhere? Or is this because of extra space for the agent to move?

[MrRobot2211]

Hi @artemZholus could you take a look at this. I am worried about the ranges in the block's coordinates.

[artemZholus]

Hi @MrRobot2211 !

To avoid confusion, I recommend you work only with the parsed representation of the dataset. Several examples of working with the data are shown here: https://github.com/iglu-contest/gridworld/tree/master/examples . The general structure of the dataset is described here: https://github.com/iglu-contest/gridworld/blob/master/README.md#working-with-the-iglu-dataset .

Regarding your question, in gridworld['worldEndingState']['blocks'] you can find a list of 4-tuples of the form: (x, y, z, block_id). This is the raw data, therefore where the coordinate origin is at the center of the building zone (which is why x and z coords can have values between -5 and 5). The block id values can have 5 values between 0 and 100 because this is the raw data from the data collector (which is why we recommend you to use the parser, it normalizes them making them follow this mapping). The overall shape of the building zone is (9, 11, 11) the coords order is (Y, X, Z).

[MrRobot2211]

Thank you . So okey this

the gridworld['worldEndingState']['blocks'] is a list of 4-tuples representing 3-d coordinates of a block and its color.
was for the old env as is this mapping then

https://iglu-contest.github.io/tasks.html

I have seen the parsed datasets. But during the evaluation ion the NLP task we are just given the worldEndingState, not even the whole tape.

So I suppose I can reuse https://github.com/iglu-contest/gridworld/blob/c9d8f4f0a38ea8aa7adf09f551c9809923e06648/gridworld/data/iglu_dataset.py#L211
for processing and getting back an interpretable representation.

Thank you.

[MrRobot2211]

Hi @artemZholus thank you for the clarifications I have managed to make progress but I am still a little bit unsure about coordinates.

I have a particular example.

Instruction = 'Facing north destroy the purple block on the left. Place two blue blocks between the purple deleted block and the other purple blocks. They will be laying horizontally.'

blocks = [[-3, 63, 0, 56], [0, 63, -2, 56], [0, 63, -1, 56], [0, 63, 0, 56]]
pos = [-2.976793981194584, 63.47999999999999, 4.902476381694842]
obs = [-2.976793981194584, 63.47999999999999, 4.902476381694842]
look = [-0.5440000000000005, 6.225999999999974]

I painted the observer red. And I am not plotting coordinate-1 (assume 0 indexing) which references the height

My interpretation of this is that north south in the graph is given by the Y axis in the plot which correspondes to coorinate-2 of the observer and the blocks. In particular the North would be in negative Y. And east-west is axis x (coordinate-0).

But now if the observer faces north , the purple block that is alone is slightly to the right of the observer (maybe depending on how we take into account the position of the corners of the blocks).

Is my analysis correct?