Datasets:

MasoudJTehrani
/

NSAP-artifacts

Tasks:

Robotics

Tags:

adversarial-attacks

scene-graph-generation

embodied-ai

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

test

The dataset viewer is not available for this split.

Cannot load the dataset split (in streaming mode) to extract the first rows.

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
patch: string
attack_mode: string
target: string
n_test_clips_requested: int64
n_test_clips_loaded: int64
total_frames: int64
flip_rate: double
n_flipped: int64
n_total: int64
n_clips_with_targets: int64
total_target_objects: int64
total_category_changes: int64
total_attribute_changes: int64
total_relation_changes: int64
total_disappeared: int64
total_changes: int64
pooled_isolation_score: double
mean_n_target_objects: double
mean_category_change_rate: double
mean_attribute_change_rate: double
mean_relation_change_rate: double
mean_disappear_rate: double
mean_isolation_score: double
n_frames_requested: int64
_fields: struct<attack_mode: string, target: string, n_frames_requested: string, n_frames_loaded: string, n_f (... 321 chars omitted)
  child 0, attack_mode: string
  child 1, target: string
  child 2, n_frames_requested: string
  child 3, n_frames_loaded: string
  child 4, n_frames_with_clean_target: string
  child 5, flip_rate: string
  child 6, n_total: string
  child 7, n_flipped: string
  child 8, n_objects_patched: string
  child 9, n_target_relations: string
  child 10, total_category_changes: string
  child 11, total_attribute_changes: string
  child 12, total_relation_changes: string
  child 13, total_disappeared: string
  child 14, total_changes: string
  child 15, pooled_isolation_score: string
n_objects_patched: int64
n_target_relations: int64
n_frames_with_clean_target: int64
n_frames_loaded: int64
to
{'_fields': {'attack_mode': Value('string'), 'target': Value('string'), 'n_frames_requested': Value('string'), 'n_frames_loaded': Value('string'), 'n_frames_with_clean_target': Value('string'), 'flip_rate': Value('string'), 'n_total': Value('string'), 'n_flipped': Value('string'), 'n_objects_patched': Value('string'), 'n_target_relations': Value('string'), 'total_category_changes': Value('string'), 'total_attribute_changes': Value('string'), 'total_relation_changes': Value('string'), 'total_disappeared': Value('string'), 'total_changes': Value('string'), 'pooled_isolation_score': Value('string')}, 'patch': Value('string'), 'attack_mode': Value('string'), 'target': Value('string'), 'n_frames_requested': Value('int64'), 'n_frames_loaded': Value('int64'), 'n_frames_with_clean_target': Value('int64'), 'flip_rate': Value('float64'), 'n_flipped': Value('int64'), 'n_total': Value('int64'), 'n_objects_patched': Value('int64'), 'n_target_relations': Value('int64'), 'total_category_changes': Value('int64'), 'total_attribute_changes': Value('int64'), 'total_relation_changes': Value('int64'), 'total_disappeared': Value('int64'), 'total_changes': Value('int64'), 'pooled_isolation_score': Value('float64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              patch: string
              attack_mode: string
              target: string
              n_test_clips_requested: int64
              n_test_clips_loaded: int64
              total_frames: int64
              flip_rate: double
              n_flipped: int64
              n_total: int64
              n_clips_with_targets: int64
              total_target_objects: int64
              total_category_changes: int64
              total_attribute_changes: int64
              total_relation_changes: int64
              total_disappeared: int64
              total_changes: int64
              pooled_isolation_score: double
              mean_n_target_objects: double
              mean_category_change_rate: double
              mean_attribute_change_rate: double
              mean_relation_change_rate: double
              mean_disappear_rate: double
              mean_isolation_score: double
              n_frames_requested: int64
              _fields: struct<attack_mode: string, target: string, n_frames_requested: string, n_frames_loaded: string, n_f (... 321 chars omitted)
                child 0, attack_mode: string
                child 1, target: string
                child 2, n_frames_requested: string
                child 3, n_frames_loaded: string
                child 4, n_frames_with_clean_target: string
                child 5, flip_rate: string
                child 6, n_total: string
                child 7, n_flipped: string
                child 8, n_objects_patched: string
                child 9, n_target_relations: string
                child 10, total_category_changes: string
                child 11, total_attribute_changes: string
                child 12, total_relation_changes: string
                child 13, total_disappeared: string
                child 14, total_changes: string
                child 15, pooled_isolation_score: string
              n_objects_patched: int64
              n_target_relations: int64
              n_frames_with_clean_target: int64
              n_frames_loaded: int64
              to
              {'_fields': {'attack_mode': Value('string'), 'target': Value('string'), 'n_frames_requested': Value('string'), 'n_frames_loaded': Value('string'), 'n_frames_with_clean_target': Value('string'), 'flip_rate': Value('string'), 'n_total': Value('string'), 'n_flipped': Value('string'), 'n_objects_patched': Value('string'), 'n_target_relations': Value('string'), 'total_category_changes': Value('string'), 'total_attribute_changes': Value('string'), 'total_relation_changes': Value('string'), 'total_disappeared': Value('string'), 'total_changes': Value('string'), 'pooled_isolation_score': Value('string')}, 'patch': Value('string'), 'attack_mode': Value('string'), 'target': Value('string'), 'n_frames_requested': Value('int64'), 'n_frames_loaded': Value('int64'), 'n_frames_with_clean_target': Value('int64'), 'flip_rate': Value('float64'), 'n_flipped': Value('int64'), 'n_total': Value('int64'), 'n_objects_patched': Value('int64'), 'n_target_relations': Value('int64'), 'total_category_changes': Value('int64'), 'total_attribute_changes': Value('int64'), 'total_relation_changes': Value('int64'), 'total_disappeared': Value('int64'), 'total_changes': Value('int64'), 'pooled_isolation_score': Value('float64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

NSAP artifacts

Trained patches and benchmark run logs for the NSAP paper. These are the files a replicator cannot regenerate: retraining gives different patches, and re-running the embodied benchmark gives different trajectories, because gpt-4o is sampled.

Everything else (VINE weights, GroundingDINO, SAM2, the LASER source, the simulated dataset) is fetched by python NSAP.py --with-detection --with-dataset in the code repository.

Use

This repo mirrors the code repository's output/ directory, so:

bash replicate.sh fetch-artifacts   # unpacks straight into output/
bash replicate.sh eval-model        # reproduces every model-level number

universal_patches/<head>/                       Stage 1, one patch per head (5)
chameleon_patches/<head>/<scenario>/<head>/     Stage 2, per-scenario (30)
    *_final.pt            the patch: weights, anchor, and the full AttackConfig it trained under
    *_final_eval.json     its model-level evaluation
    *_final_preview.png   the patch alone, and composited on its target

rung2/<run>/<rep>/<condition>/<episode>/
    steps.jsonl           per step: the scene graph, the text the planner received,
                          its full response, the action, patch state, distances
    trace.txt             the same, rendered for reading
    run_meta.json         scene, instruction, patch kind, aggr_thres, model, start time
rung2/<run>/<rep>/<condition>/results.json      episode outcomes
rung2/<run>/summary.json                        the aggregate for that run

outcome_eval.json       per-head flip rate, isolation, and the outcome taxonomy
deploy_eval.json        deployment-faithful scoring through predict_gd -> predict_laser
deploy_eval_topk15.json the same with the graph widened to 15 objects
deploy_eval_random.json the random-patch control
stealth_eval.json       LPIPS, size-matched, per viewpoint

The run directories

Each is one session, and the name records the configuration, which changes the answer:

run	configuration
`staged_A_donothing_tau0.1`	attribute and relation tasks, `aggr_thres` 0.1, `top_k` 1
`staged_A_donothing_tau0.1_k1`	the same configuration, run again
`staged_A_donothing_tau0.1_k1_lang`	adds the language-only conditions
`staged_A_donothing_tau0.4`	the harness gate
`staged_A_donothing_tau0.4_k15`	graph widened to 15 objects
`staged_E_single_pot_tau0.4`	single-pot staging variant
`ebnav_tau0.4`	category benchmark, our configuration
`ebnav_tau0.3_faithful1`	category benchmark, ESCA's nav evaluator config (`gd_only=True`)
`ebnav_tau0.3_faithful1_gdonly0`	the same but with SGClip in the loop

The first two rows are deliberate. They share a configuration, and the spread between them is the paper's evidence on benchmark drift.

Reading the logs

python rung2/analyze_steps.py output/rung2/<run>   # diff esca_benign against esca_patch, step by step

Before trusting any single episode, check frames_with_patch. If it is 0 the target was never visible, the patch never injected, and the outcome is not attributable to the attack.

Patch checkpoints from partway through training are not included, only *_final.

Downloads last month: 44

Total file size:

224 MB

NSAP artifacts

Use

Contents

The run directories

Reading the logs