SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2105.04556 · IJCAI · 2021

TANGO: Commonsense Generalization in Predicting Tool Interactions for Mobile Manipulators

Shreshth Tuli, Rohan Paul, Rajas Bansal

arXiv · PDF · Open in the Atlas

Code that ran

8 of 14 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 7 failures of our instrument; 6 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2105.04556.json · badge: /api/ran/2105.04556.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
reail-iitd/tango canonical (classifier label)
classifier basis: census
8 ran · 6 no recorded run
FunctionStatusWhere it lives
convertActions Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/src/actions.py
code served (permissive licence) · get_code("5842c92e5964be7c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# The function convertActions takes inp (a dict with 'actions' key) and world (a string or dict)
# It iterates through high_level_actions, each with 'name' and 'args' keys
# The world parameter is checked with 'in' operator (string containment check)

torch.manual_seed(0)

# Construct a minimal inp dictionary with various action types to test the function
inp = {
    'actions': [
        {
            'name': 'pickNplaceAonB',
            'args': ['obj1', 'obj2']
        },
        {
            'name': 'changeWing',
            'args': ['down']
        },
        {
            'name': 'moveAToB',
            'args': ['loc1', 'loc2']
        },
        {
            'name': 'push',
            'args': ['obj1', [1.0, 2.0, 3.0]]
        },
        {
            'name': 'pushTo',
            'args': ['obj1', [4.0, 5.0]]
        },
        {
            'name': 'moveTo',
            'args': ['location1']
        },
        {
            'name': 'move',
            'args': [[0.1, 0.2, 0.3]]
        },
        {
            'name': 'moveUp',
            'args': []
        },
        {
            'name': 'moveDown',
            'args': []
        },
        {
            'name': 'pick',
            'args': ['obj1']
        },
        {
            'name': 'place',
            'args': ['obj1', 'loc1']
        },
        {
            'name': 'dropTo',
            'args': ['obj1', 'loc1']
        },
        {
            'name': 'drop',
            'args': ['obj1']
        },
        {
            'name': 'changeState',
            'args': ['obj1', 'state1']
        },
        {
            'name': 'placeRamp',
            'args': []
        },
        {
            'name': 'stick',
            'args': ['obj1', 'loc1']
        },
        {
            'name': 'fuel',
            'args': ['loc1', 'fuel_item']
        },
        {
            'name': 'cut',
            'args': ['loc1', 'cutter']
        },
        {
            'name': 'print',
            'args': ['model1']
        },
        {
            'name': 'drive',
            'args': ['obj1', 'tool1']
        },
        {
            'name': 'weld',
            'args': ['obj1']
        },
        {
            'name': 'paint',
            'args': ['obj1']
        },
        {
            'name': 'drill',
            'args': ['obj1']
        },
        {
            'name': 'apply',
            'args': ['obj1', 'obj2']
        },
        {
            'name': 'climbUp',
            'args': ['ladder1']
        },
        {
            'name': 'climbDown',
            'args': ['ladder1']
        },
        {
            'name': 'clean',
            'args': ['area1']
        }
    ]
}

# world is checked with string containment ('in' operator)
# Test with 'home' to trigger the home branch in stick action
world = 'home_warehouse'

ARGS = {
    'inp': inp,
    'world': world
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convertActionsFromFile Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/src/actions.py
code served (permissive licence) · get_code("6f8eb99c754ceedf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary JSON file with sample action data
temp_dir = tempfile.mkdtemp()
action_file = os.path.join(temp_dir, "actions.json")

# Write sample action data to the file
sample_actions = {
    "actions": [
        {"id": 0, "type": "move", "target": "left"},
        {"id": 1, "type": "move", "target": "right"},
        {"id": 2, "type": "jump", "target": "up"}
    ],
    "metadata": {
        "version": "1.0",
        "count": 3
    }
}

with open(action_file, 'w') as f:
    json.dump(sample_actions, f)

ARGS = {
    "action_file": action_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
form_goal_vec Ran
ran, honoured the contract we drafted
reail-iitd/tango/src/extract_vectors.py
code served (permissive licence) · get_code("e4f260449b2520a6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (300,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a small embedding data dictionary
# Each word maps to a 300-dimensional vector
data = {
    "hello": np.random.randn(300),
    "world": np.random.randn(300),
    "test": np.random.randn(300),
    "example": np.random.randn(300),
}

# Create a simple text string with words that exist in data
text = "hello world test"

ARGS = {
    "data": data,
    "text": text,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (300,) — shape (300,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
getInteractedObjs Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/analyze.py
code served (permissive licence) · get_code("21aba124f362dee5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary directory and a mock objects.json file
temp_dir = tempfile.mkdtemp()
jsons_dir = os.path.join(temp_dir, 'jsons')
os.makedirs(jsons_dir, exist_ok=True)

# Write a mock objects.json file
objects_data = {
    'objects': [
        {'name': 'apple'},
        {'name': 'banana'},
        {'name': 'orange'},
        {'name': 'table'},
        {'name': 'chair'}
    ]
}

objects_file_path = os.path.join(jsons_dir, 'objects.json')
with open(objects_file_path, 'w') as f:
    json.dump(objects_data, f)

# Change to temp directory so the relative path works
original_dir = os.getcwd()
os.chdir(temp_dir)

# Create a minimal datapoint class with symbolicActions attribute
class MockDatapoint:
    def __init__(self, symbolic_actions):
        self.symbolicActions = symbolic_actions

# Create mock actions
# Each action is a tuple where action[0] is a dict with 'args' key
action1 = ({'args': ['apple', 'table', 'banana']},)
action2 = ({'args': ['chair', 'orange', 'apple']},)
action3 = ('E',)  # Should be skipped (starts with 'E')
action4 = ('U',)  # Should be skipped (starts with 'U')
action5 = ({'args': ['nonexistent', 'table', 'xyz']},)  # Mixed valid/invalid

datapoint = MockDatapoint([action1, action2, action3, action4, action5])

ARGS = {
    'datapoint': datapoint
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
initWingPos Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/src/initialise.py
code served (permissive licence) · get_code("8184f2361f9ea6d0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary JSON file with valid wing pose data
temp_dir = tempfile.gettempdir()
wing_file_path = os.path.join(temp_dir, "wing_poses_test.json")

# Create the JSON data structure expected by the function
wing_data = {
    "poses": [
        {
            "name": "home",
            "pose": [0.0, -1.57, 1.57, -1.57, -1.57, 0.0, 0.0]
        },
        {
            "name": "ready",
            "pose": [0.5, -1.2, 1.2, -1.2, -1.57, 0.5, 0.1]
        },
        {
            "name": "grab",
            "pose": [1.0, -0.8, 0.8, -0.8, -1.57, 1.0, 0.8]
        }
    ]
}

# Write the JSON file
with open(wing_file_path, 'w') as f:
    json.dump(wing_data, f)

ARGS = {
    "wing_file": wing_file_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
listSum Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/src/generalization.py
code served (permissive licence) · get_code("bf901dc21c7d331e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Analyze the function:
# - listSum takes two parameters: a and b
# - It iterates over range(len(a)), so both a and b must have a len()
# - It accesses a[i] and b[i] with indexing, so they must be indexable
# - It adds a[i] + b[i] and rounds to 4 decimal places
# - Returns a list of rounded sums

# Create simple list arguments
a = [1.23456, 2.34567, 3.45678]
b = [4.56789, 5.67890, 6.78901]

ARGS = {
    "a": a,
    "b": b
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_all_vectors Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/src/extract_vectors.py
code served (permissive licence) · get_code("34f8e4041ae2e7c5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import io
import tempfile
import os

# Create a temporary file with word vectors in fastText format
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.vec', encoding='utf-8')
temp_filename = temp_file.name

try:
    # Write header: number of words and dimension
    temp_file.write("3 5\n")  # 3 words, 5 dimensions
    
    # Write word vectors
    temp_file.write("apple 0.1 0.2 0.3 0.4 0.5\n")
    temp_file.write("banana 0.2 0.3 0.4 0.5 0.6\n")
    temp_file.write("cherry 0.3 0.4 0.5 0.6 0.7\n")
    
    temp_file.close()
    
    ARGS = {
        'fname': temp_filename
    }
finally:
    pass
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
world_split Ran
ran; our contract draft was wrong, not the code
reail-iitd/tango/rl.py
code served (permissive licence) · get_code("30e7a00d1c9c5573")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

class MockGraphData:
    def __init__(self):
        self.graphs = [
            (1, 1),
            (2, 2),
            (3, 4),
            (5, 5),
            (6, 7),
            (8, 8),
            (9, 10),
        ]

data = MockGraphData()

ARGS = {
    "data": data
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
allActionTypes Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/analyze.py
code served (permissive licence) · get_code("cff902e62a7a6b18")
How it ran

no invocation recorded.

fct Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/approx.py
code served (permissive licence) · get_code("014dfa205be068ec")
How it ran

no invocation recorded.

fcw Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/approx.py
code served (permissive licence) · get_code("2ca7da4ad6dccbc3")
How it ran

no invocation recorded.

getInterestTools Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/analyze.py
code served (permissive licence) · get_code("1f193210eddf07cd")
How it ran

no invocation recorded.

load_vectors Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/src/extract_vectors.py
code served (permissive licence) · get_code("b52b00285d990bde")
How it ran

no invocation recorded.

load_world Not yet run
unverified (harvested, no recorded run)
reail-iitd/tango/approx.py
code served (permissive licence) · get_code("a32d700177f2271e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 14 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Robots assisting us in factories or homes must learn to make use of objects as tools to perform tasks, e.g., a tray for carrying objects. We consider the problem of learning commonsense knowledge of when a tool may be useful and how its use may be composed with other tools to accomplish a highlevel task instructed by a human. We introduce a novel neural model, termed TANGO, for predicting task-specific tool interactions, trained using demonstrations from human teachers instructing a virtual robot in a physics simulator. TANGO encodes the world state, comprising objects and symbolic relationships between them, using a graph neural network. The model learns to attend over the scene using knowledge of the goal and the action history, finally decoding the symbolic action to execute. Crucially, we address generalization to unseen environments where some known tools are missing, but alternative unseen tools are present. We show that by augmenting the representation of the environment with pre-trained embeddings derived from a knowledge-base, the model can generalize effectively to novel environments. Experimental results show a 60.5-78.9% absolute improvement over the baseline in predicting successful symbolic plans in unseen settings for a simulated mobile manipulator.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2105.04556")
get_code_for_paper("2105.04556")
have("2105.04556")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2105.04556.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2105.04556.svg)](https://syntology.ai/paper/2105.04556)

Connect an agent — have() is free.