ArraMon

ArraMon: A Joint Navigation-Assembly Instruction Interpretation Task in Dynamic Environments

Hyounghun Kim, Abhay Zala, Graham Burri, Hao Tan, Mohit Bansal

What is ArraMon / ऐरेमौन?

ArraMon (PokeMon + Arrangement) is a new task/dataset for learning navigation+assembling guided by natural language instructions. The agents in the ArraMon task should develop integrated abilities to map directional phrases in the instruction to actions in large realistic outdoor environments and grid-layout environments (both 3D egocentric), as well as understand object referring and spatial/depth expressions to perform the task successfully.

Paper / Code / Dataset

Please see our EMNLP Findings Paper.

Paper

Please see our Github page for the code/dataset and how to get started.

Github

We have also added Hindi instructions (val seen and unseen splits released, training+test splits coming soon).

Citation

If you use ArraMon please cite our paper with this BibTeX entry.


								@inproceedings{Kim2020ArraMonAJ,

									  title={ArraMon: A Joint Navigation-Assembly Instruction Interpretation Task in Dynamic Environments},

									  author={Hyounghun Kim and Abhaysinh Zala and Graham Burri and Hao Tan and Mohit Bansal},

									  booktitle={Findings of EMNLP},

									  year={2020}

								}

Score List

Please contact arramonunc@gmail.com for the test-unseen split.

Rank	Model	Code	nDTW	CTC				rPOD	PTC
Rank	Model	Code	nDTW	k=0	k=3	k=5	k=7	rPOD	PTC
1 June 1st, 2020	Baseline Model UNC Chapel Hill (Kim, Zala, et al. 2020)	Github	0.114	0.000	0.082	0.122	0.168	0.047	0.035