Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand


  • Description:

Grounded SCAN (gSCAN) is a synthetic dataset for evaluating compositional generalization in situated language understanding. gSCAN pairs natural language instructions with action sequences, and requires the agent to interpret instructions within the context of a grid-based visual navigation environment.

More information can be found at:

    'command': Sequence(Text(shape=(), dtype=string)),
    'manner': Text(shape=(), dtype=string),
    'meaning': Sequence(Text(shape=(), dtype=string)),
    'referred_target': Text(shape=(), dtype=string),
    'situation': FeaturesDict({
        'agent_direction': int32,
        'agent_position': FeaturesDict({
            'column': int32,
            'row': int32,
        'direction_to_target': Text(shape=(), dtype=string),
        'distance_to_target': int32,
        'grid_size': int32,
        'placed_objects': Sequence({
            'object': FeaturesDict({
                'color': Text(shape=(), dtype=string),
                'shape': Text(shape=(), dtype=string),
                'size': int32,
            'position': FeaturesDict({
                'column': int32,
                'row': int32,
            'vector': Text(shape=(), dtype=string),
        'target_object': FeaturesDict({
            'object': FeaturesDict({
                'color': Text(shape=(), dtype=string),
                'shape': Text(shape=(), dtype=string),
                'size': int32,
            'position': FeaturesDict({
                'column': int32,
                'row': int32,
            'vector': Text(shape=(), dtype=string),
    'target_commands': Sequence(Text(shape=(), dtype=string)),
    'verb_in_command': Text(shape=(), dtype=string),
  • Feature documentation:
Feature Class Shape Dtype Description
command Sequence(Text) (None,) string
manner Text string
meaning Sequence(Text) (None,) string
referred_target Text string
situation FeaturesDict
situation/agent_direction Tensor int32
situation/agent_position FeaturesDict
situation/agent_position/column Tensor int32
situation/agent_position/row Tensor int32
situation/direction_to_target Text string
situation/distance_to_target Tensor int32
situation/grid_size Tensor int32
situation/placed_objects Sequence
situation/placed_objects/object FeaturesDict
situation/placed_objects/object/color Text string
situation/placed_objects/object/shape Text string
situation/placed_objects/object/size Tensor int32
situation/placed_objects/position FeaturesDict
situation/placed_objects/position/column Tensor int32
situation/placed_objects/position/row Tensor int32
situation/placed_objects/vector Text string
situation/target_object FeaturesDict
situation/target_object/object FeaturesDict
situation/target_object/object/color Text string
situation/target_object/object/shape Text string
situation/target_object/object/size Tensor int32
situation/target_object/position FeaturesDict
situation/target_object/position/column Tensor int32
situation/target_object/position/row Tensor int32
situation/target_object/vector Text string
target_commands Sequence(Text) (None,) string
verb_in_command Text string
 author = {Ruis, Laura and Andreas, Jacob and Baroni, Marco and Bouchacourt, Diane and Lake, Brenden M},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {H. Larochelle and M. Ranzato and R. Hadsell and M. F. Balcan and H. Lin},
 pages = {19861--19872},
 publisher = {Curran Associates, Inc.},
 title = {A Benchmark for Systematic Generalization in Grounded Language Understanding},
 url = {},
 volume = {33},
 year = {2020}

    title = "Systematic Generalization on g{SCAN}: {W}hat is Nearly Solved and What is Next?",
    author = "Qiu, Linlu  and
      Hu, Hexiang  and
      Zhang, Bowen  and
      Shaw, Peter  and
      Sha, Fei",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "",
    doi = "10.18653/v1/2021.emnlp-main.166",
    pages = "2180--2188",

grounded_scan/compositional_splits (default config)

  • Config description: Examples for compositional generalization.

  • Download size: 82.10 MiB

  • Dataset size: 998.11 MiB

  • Splits:

Split Examples
'adverb_1' 112,880
'adverb_2' 38,582
'contextual' 11,460
'dev' 3,716
'situational_1' 88,642
'situational_2' 16,808
'test' 19,282
'train' 367,933
'visual' 37,436
'visual_easier' 18,718


  • Config description: Examples for generalizing to larger target lengths.

  • Download size: 53.41 MiB

  • Dataset size: 546.73 MiB

  • Splits:

Split Examples
'dev' 1,821
'target_lengths' 198,588
'test' 37,784
'train' 180,301


  • Config description: Examples for spatial relation reasoning.

  • Download size: 89.59 MiB

  • Dataset size: 675.09 MiB

  • Splits:

Split Examples
'dev' 2,617
'referent' 30,492
'relation' 6,285
'relative_position_1' 41,576
'relative_position_2' 41,529
'test' 28,526
'train' 259,088
'visual' 62,250