Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Grounded SCAN

This repository contains the code for generating the grounded SCAN benchmark. Grounded SCAN poses a simple task, where an agent must execute action sequences based on a synthetic language instruction.

The agent is presented with a simple grid world containing a collection of objects, each of which is associated with a vector of features. The agent is evaluated on its ability to follow one or more instructions in this environment. Some instructions require interaction with particular kinds of objects.

NB: for reinforcement learning mode see branch rlmode. Please note that the full reward function including interactions and manners has not been tested with a learned policy yet.


The data used in the grounded SCAN paper can be found in the folder data of this repository. This data can be used to train models with the multi-modal baseline from the paper. The exact match accuracies reported in the paper can then be reproduced with the mode error_analysis in this repository.

Some data examples (more at bottom of this file, and a demo dataset can be found in data/demo_dataset/.):

Random Example while zigzagging


This section contains the leaderboard for scores obtained by papers on gSCAN. To add scores please consider a pull request.

Baseline GECA [1] [2] [3] [4] [5] [6]
A: Random 97.69 +- 0.22 87.6 +- 1.19 97.32 94.19 +- 0.71 98.6 +- 0.95 - 74.7 99.95 +/- 0.02
B: Yellow Squares 54.96 +- 39.39 34.92 +- 39.30 95.35 87.31 +- 4.38 99.08 +- 0.69 - 81.3 99.90 +/- 0.06
C: Red Squares 23.51 +- 21.82 78.77 +- 6.63 80.16 81.07 +- 10.12 80.31 +- 24.51 - 78.1 99.25 +/- 0.91
D: Novel Direction 0.00 +- 0.00 0.00 +- 0.00 5.73 - 0.16 +- 0.12 - 0 0.0 +/- 0.0
E: Relativity 35.02 +- 2.35 33.19 +- 3.69 75.19 52.8 +- 9.96 87.32 +- 27.38 - 53.6 99.02 +/- 1.16
F: Class Inference 92.52 +- 6.75 85.99 +- 0.85 98.63 - 99.33 +- 0.46 - 76.2 99.98 +/- 0.01
G: Adverb k=1 0.00 +- 0.00 0.00 +- 0.00 11.94 - - - 0.0 0.00 +/- 0.00
G: Adverb k=5 0.47 +- 0.14 - 10.31 - - - - -
G: Adverb k=10 2.04 +- 0.95 - 33.28 - - 4.87 4.87 -
G: Adverb k=50 4.63 +- 2.08 - 40.78 - - - - -
H: Adverb to Verb 22.70 +- 4.59 11.83 +- 0.31 21.95 - 33.6 +- 20.81 28.03 21.8 22.16 +/- 0.01
I: Length 2.10 +- 0.05 - - - - -

[1] Yen-Ling Kuo, Boris Katz, and Andrei Barbu. 2020. "Compositional networks enable systematic generalization for grounded language understanding."

[2] Christina Heinze-Deml and Diane Bouchacourt. 2020. "Think before you act: A simple baseline for compositional generalization." in EMNLP 2020

[3] Tong Gao, Qi Huang, Raymond J. Mooney. 2020. "Systematic Generalization on gSCAN with Language Conditioned Embedding" in AACL-IJCNLP 2020.

[4] Yichen Jiang and Mohit Bansal. 2021. "Inducing Transformer’s Compositional Generalization Ability via Auxiliary Sequence Prediction Tasks"

[5] Maxwell Nye, Michael Henry Tessler, Joshua B. Tenenbaum, Brenden M. Lake. 2021. "Improving Coherence and Consistency in Neural Sequence Models with Dual-System, Neuro-Symbolic Reasoning"

[6] Linlu Qui, Hexiang Hu, Bowen Zhang, Peter Shaw, Fei Sha. 2020. "Systematic Generalization on gSCAN: What is Nearly Solved and What is Next?"

Data for Paper

All data used in the paper can be found in compressed folders in data. To see how to read this data for a computational model and train models on it please refer to this repository. Here the code for data generation, error analysis, and visualizing examples lives.

Extra Documentation

Find a term glossary in documentation/

Getting Started

Make a virtualenvironment that uses Python 3.7 or higher:

>> virtualenv --python=/usr/bin/python3.7 <path/to/virtualenv>

Activate the environment and install the requirements with a package manager:

>> { source <path/to/virtualenv>/bin/activate; python3.7 -m pip install -r requirements; }

Make sure all tests run correctly:

(virtualenv) >> python3.7 -m GroundedScan --mode=test

Demo Dataset

In the folder data/demo_dataset a very small example dataset can be found that can be inspected to get a feel for the data. It contains an example of the compositional generalization splits you could create with a simple intransitive grammar (i.e., without size modifiers, adverbs, or transitive verbs), with a grid size of 4, and only 2 shapes. When taking a look in the folder, one can also see all the extra statistics and plots that are generated when generating a benchmark dataset. The actual data that can be used for training computational models can be found in dataset.txt. To see a summary of the statistics of each data split in dataset.txt, inspect the separate split text files named <split>_dataset_stats.txt (e.g., train_dataset_stats.txt). This demo dataset is generated by the following command:

>> python3.7 -m GroundedScan --mode=generate --output_directory=data/dummy_dataset --num_resampling=1 --grid_size=4 --type_grammar=simple_intrans --nouns=circle,square --percentage_dev=0.1 --make_dev_set

See an demonstration of training a multi-modal neural model on this demo dataset here.

Using the Repository

Generating Data

To generate data for the compositional generalization splits:

(virtualenv) >> python3.7 -m GroundedScan --mode=generate --output_directory=compositional_splits --type_grammar=adverb --split=generalization

in a directory containing the folder GroundedScan. This will generate a dataset in the folder compositional_splits. The dataset will be in compositional_splits/dataset.txt and for each split there will be an associated compositional_splits/<split>_stats.txt and plots visualizing the statistics of the data in that particular split (e.g. the number of examples totally, broken down per grid location, per referred target, etc.). For an example dataset see the folder data/, containing the data used for the paper.

Important Parameters

  • grid_size specifies the number of columns and rows in the world. If set higher, more examples will be generated and the maximum target length will be higher.
  • num_resampling this determines how often you want to resample an example with the same specifications in terms of input instruction, target referrent, and relative position of the agent versus the target. If set to higher than 1, the code samples new random agent and target locations that still satisfy the constraints. Like setting a higher grid size, this parameter can also be used to generate more data.
  • split determines for which splits we are generating data. Can be set to uniform (i.e. no systematic difference between training and test), generalization (i.e. the compositional generalization splits), or target_lengths (i.e. generalizing to larger target lengths). For this latter split make sure to also set cut_off_target_length
  • visualize_per_template is an integer determining how many gifs to make per template example (see for explanation of template).
  • type_grammar if set to adverb just gives the fully implemented grammar from the paper, but you can also set it to normal to exclude adverbs, or to simple_intrans and simple_trans to use only intransitive or transitive verbs respectively.

Error Analysis

To do an error analysis over a predictions file generated by a model trained on the data, run:

(virtualenv) >> python3.7 -m GroundedScan --mode=error_analysis --output_directory=error_analysis --load_dataset_from=compositional_splits/dataset.txt  --predicted_commands_file=predict.json

The predicted commands file must be a file generated with a trained model in test mode in the provided model code. The file example_prediction.json contains 1 data example with a prediction that can be used for trying out the error analysis and execute commands modes. The error analysis will generate plots in the output directory specified, as well as a .txt and .csv file summarizing all the results for own anaylsis. The .csv outputs were used in the paper.

NB: this mode is used to get the exact match accuracies reported in the paper.

Execute Commands

To make gifs of predictions use the mode execute_commands:

(virtualenv) >> python3.7 -m GroundedScan --mode=execute_commands --load_dataset_from=compositional_splits/dataset.txt --output_directory=visualized_examples

This will visualize the predictions in predict.json that should be placed in the folder specified by --output_directory and visualize the execution in a gif in the specified output directory. NB: this will make one gif (including as many images as there are time-steps in that prediction) for each data point in predict.json. If you want to only inspect the errors, set --only_save_errors. The file example_prediction.json contains 1 data example with a prediction that can be used for trying out the error analysis and execute commands modes.

Implementation Details

We construct a Grammar that produces transitive and intransitive sentences (with a depth-limited recursive operation for introducing adjectives and adverbs). Sentences are of the form

((Adverb* Verb (Adjective* Noun)?)+

It is possible to sample from a Grammar to obtain paired sentences (sequences of tokens) and meanings (hacky neo-Davidsonian logical forms). For example, the sentence

push a big square

gets associated with the logical form

lambda $v1. exists $n. push($v1) and patient($v1, $n) and big($n) and

represented in the code by

(push x0:verb) ^ (big x2:noun) ^ (square x2:noun) ^ (patient x0:verb x2:noun)

In order to pair instructions with demonstrations, we need to map from meanings to sequences of low-level environment actions.


The example logical form depicted above specifies one action: a square-pushing action. It must be performed in a grounded context. In this library, computations associated with grounding live in The grounded context for a specific instruction is represented by a Situation. Situations are generated from Worlds. A situation is what is used for one example to specify all the important information to generate a visualization as well as input grid world for a computational model. The function to visualize a current situation in an RGB image is World.get_current_situation_image() in and to get the simplified input to a model (see paper Section 3 (World Model) for explanation) run the function World.get_current_situation_grid_repr(). This is all done automatically when generating the data.

Important semantics:

  • walk: means walking somewhere with action command walk, and for navigation turn left and turn right.
  • push: means pushing an object to the front of the agent until it hits a wall or another object, to move it 1 grid cell takes 1 push action if the object is of size 1 or 2, and 2 push actions push push if it is of size 3 or 4.
  • pull: means pulling an object to the back of the agent until it hits a wall or another object, with pull-actions, again needing 1 action for light objects (size 1 and 2), and 2 for heavy ones (size 3 and 4).
  • while spinning: generate a sequence of turn left turn left turn left turn left (i.e. one spin) every time before executing the actions to move a grid cell.
  • cautiously: generate a sequence of turn right turn left turn left turn right (i.e. look to the right and left) each time before you move over a grid line (note the distinction with when to execute the sequence from while spinning, see examples at bottom of this page for clarification).
  • hesitantly: generate a action command to stay every time after moving a grid cell.
  • while zigzagging: instead of walking all the way horizontally first and then vertically, alternate between horizontal and vertical until in line with the target, from then go straight.

Code-Flow Diagram

For a schematic depiction of how exactly the code is setup see the following image: Code flow


After instructions have been generated, we construct various held-out sets designed to test different notions of compositional generalization. This is done in the class method GroundedScan.assign_splits() in GroundedScan/

object properties

  • visual: all examples where red squares are the target object.
  • visual_easier: all examples where yellow squares are referred to with a color and a shape at least.


  • situational_1: all examples where the target object is to the South-West of the agent.
  • situational_2: all examples where the target object is a circle of size 2 and is being referred to with the small modifier.


  • contextual: all examples where the agent needs to push a square of size 3.


  • adverb_1: all examples with the adverb 'cautiously', of which we randomly select k to go in the training set.
  • adverb_2: all examples with the adverb 'while spinning' and the verb 'pull'.


  • target_lengths: all examples with a target length above --cut_off_target_length

Example gridworld command and demonstration

Compositional Splits Examples

Visual (Red Squares)

Red Squares

Visual easier (Yellow Squares)

Yellow Squares

Situational 1 (Novel Direction)

Novel Direction

Situational 2 (Relativity)

Relativity Size

Contextual (Class Inference)

Latent Class Inference

Adverb 1 (Adverb cautiously k-shot)

Adverb 1

Adverb 2 (Adverb while spinning to verb pull)

Adverb 2

Adverb Examples

While Spinning

while spinning





While Zigzagging

while zigzagging


Grounded SCAN data set.







No releases published


No packages published