Datasets:

AD-Assurance-Lab
/

steering-verification-captures

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
certificate: string
derived_by: string
note: string
lap_end_m: double
capture_span_m: struct<lap_eastbound_clear.npz: double, lap_eastbound_fog.npz: double, lap_eastbound_night.npz: doub (... 170 chars omitted)
  child 0, lap_eastbound_clear.npz: double
  child 1, lap_eastbound_fog.npz: double
  child 2, lap_eastbound_night.npz: double
  child 3, lap_eastbound_shadows.npz: double
  child 4, lap_westbound_clear.npz: double
  child 5, lap_westbound_fog.npz: double
  child 6, lap_westbound_night.npz: double
  child 7, lap_westbound_shadows.npz: double
capture_requested_m: struct<lap_eastbound_clear.npz: double, lap_eastbound_fog.npz: double, lap_eastbound_night.npz: doub (... 170 chars omitted)
  child 0, lap_eastbound_clear.npz: double
  child 1, lap_eastbound_fog.npz: double
  child 2, lap_eastbound_night.npz: double
  child 3, lap_eastbound_shadows.npz: double
  child 4, lap_westbound_clear.npz: double
  child 5, lap_westbound_fog.npz: double
  child 6, lap_westbound_night.npz: double
  child 7, lap_westbound_shadows.npz: double
capture_sha256_16: struct<lap_eastbound_clear.npz: string, lap_eastbound_fog.npz: string, lap_eastbound_night.npz: stri (... 170 chars omitted)
  child 0, lap_eastbound_clear.npz: string
  child 1, lap_eastbound_fog.npz: string
  child 2, lap_eastbound_night.npz: string
  child 3, lap_eastbound_shadows.npz: string
  child 4, lap_westbound_clear.npz: string
  child 5, lap_westbound_fog.npz: string
  child 6, lap_westbound_night.npz: string
  child 7, lap_westbound_shadows.npz: string
consistent: bool
threshold: double
cells_expected: int64
cells: list<item: struct<student: string, checkpoint: string, direction: string, cond: string, mean_abs_dif (... 39 chars omitted)
  child 0, item: struct<student: string, checkpoint: string, direction: string, cond: string, mean_abs_diff: double,  (... 27 chars omitted)
      child 0, student: string
      child 1, checkpoint: string
      child 2, direction: string
      child 3, cond: string
      child 4, mean_abs_diff: double
      child 5, poses: int64
      child 6, passed: bool
worst: double
to
{'threshold': Value('float64'), 'worst': Value('float64'), 'cells': List({'student': Value('string'), 'checkpoint': Value('string'), 'direction': Value('string'), 'cond': Value('string'), 'mean_abs_diff': Value('float64'), 'poses': Value('int64'), 'passed': Value('bool')}), 'cells_expected': Value('int64')}
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
              certificate: string
              derived_by: string
              note: string
              lap_end_m: double
              capture_span_m: struct<lap_eastbound_clear.npz: double, lap_eastbound_fog.npz: double, lap_eastbound_night.npz: doub (... 170 chars omitted)
                child 0, lap_eastbound_clear.npz: double
                child 1, lap_eastbound_fog.npz: double
                child 2, lap_eastbound_night.npz: double
                child 3, lap_eastbound_shadows.npz: double
                child 4, lap_westbound_clear.npz: double
                child 5, lap_westbound_fog.npz: double
                child 6, lap_westbound_night.npz: double
                child 7, lap_westbound_shadows.npz: double
              capture_requested_m: struct<lap_eastbound_clear.npz: double, lap_eastbound_fog.npz: double, lap_eastbound_night.npz: doub (... 170 chars omitted)
                child 0, lap_eastbound_clear.npz: double
                child 1, lap_eastbound_fog.npz: double
                child 2, lap_eastbound_night.npz: double
                child 3, lap_eastbound_shadows.npz: double
                child 4, lap_westbound_clear.npz: double
                child 5, lap_westbound_fog.npz: double
                child 6, lap_westbound_night.npz: double
                child 7, lap_westbound_shadows.npz: double
              capture_sha256_16: struct<lap_eastbound_clear.npz: string, lap_eastbound_fog.npz: string, lap_eastbound_night.npz: stri (... 170 chars omitted)
                child 0, lap_eastbound_clear.npz: string
                child 1, lap_eastbound_fog.npz: string
                child 2, lap_eastbound_night.npz: string
                child 3, lap_eastbound_shadows.npz: string
                child 4, lap_westbound_clear.npz: string
                child 5, lap_westbound_fog.npz: string
                child 6, lap_westbound_night.npz: string
                child 7, lap_westbound_shadows.npz: string
              consistent: bool
              threshold: double
              cells_expected: int64
              cells: list<item: struct<student: string, checkpoint: string, direction: string, cond: string, mean_abs_dif (... 39 chars omitted)
                child 0, item: struct<student: string, checkpoint: string, direction: string, cond: string, mean_abs_diff: double,  (... 27 chars omitted)
                    child 0, student: string
                    child 1, checkpoint: string
                    child 2, direction: string
                    child 3, cond: string
                    child 4, mean_abs_diff: double
                    child 5, poses: int64
                    child 6, passed: bool
              worst: double
              to
              {'threshold': Value('float64'), 'worst': Value('float64'), 'cells': List({'student': Value('string'), 'checkpoint': Value('string'), 'direction': Value('string'), 'cond': Value('string'), 'mean_abs_diff': Value('float64'), 'poses': Value('int64'), 'passed': Value('bool')}), 'cells_expected': Value('int64')}
              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.

Steering verification captures

Rendered camera frames along two driving routes, in four weather conditions each. These are the input to the formal certificates in formal-verification--steering--code, and they are published because they are what makes those certificates checkable without a simulator.

The paper is Testing Between the Test Cases: Proving End-to-End Steering in Conditions You Never Drove (arXiv:2609.10951).

	route	conditions	poses	frames
`captures/arterial/`	one 2,289 m lap, Town06	clear, fog, night, low sun	1,060	168×56
`captures/highway/`	2,988 m both directions, Town04	clear, fog, night, low sun	1,492 each	84×28

Each .npz holds one condition along one route with the pose track it was captured at.

teachers/ holds the four networks the shipped students were distilled from. They are needed only to re-distil a student without re-running data aggregation, so they live here rather than in every clone of the code: python3 scripts/fetch_captures.py --teachers.

645 MB in total.

Using it

From a clone of the code repository, one command puts every file where the certifier looks and checks each against a recorded digest:

python3 scripts/fetch_captures.py
STUDY_MAP=Town06 python3 scripts/verify/certify_town06.py --out /tmp/cert.json

Check the digests. A capture is the certifier's entire input, so a bound computed from the wrong frames is a statement about a different experiment, and it still prints a verdict and a margin and looks finished. SHA256SUMS lists all fifteen.

What reproduces

Every verdict. The bounds reproduce to about 4 parts in 1,000, not exactly: branch-and- bound makes different splitting choices when tiny floating-point differences reorder them, and that grows with network size. Every verdict has at least 309× more headroom than that, so the drift cannot change a conclusion, but a changed verdict means something real is different.

Citation

@software{ad_assurance_lab_steering_verification,
  author = {{AD Assurance Lab, Western Michigan University}},
  title  = {Formal verification of end-to-end steering under
            physically parameterized weather},
  year   = {2026},
  doi    = {10.5281/zenodo.22101297},
  url    = {https://github.com/AD-Assurance-Lab/formal-verification--steering--code}
}

Apache 2.0, the same as the code.

Downloads last month: 174

Total file size:

675 MB

Paper for AD-Assurance-Lab/steering-verification-captures

Testing Between the Test Cases: Proving End-to-End Steering in Conditions You Never Drove

Paper • 2609.10951 • Published 5 days ago