Datasets:

nur-dev
/

robot-execution-validation

license: cc-by-4.0
language:
  - en
size_categories:
  - 1K<n<10K
pretty_name: Robot Execution Validation
tags:
  - robotics
  - vision-language-action
  - robot-manipulation
  - evaluation
  - reproducibility

Robot Execution Validation

Execution and instruction-control records for released SmolVLA and pi0 policies in VLABench. The live validation contains 1860 scheduled attempts: 420 native controls and 1440 paired-scene rollouts, using inference seeds 11, 29 and 47. The 60 nominal paired fixtures represent 44 distinct initial scenes. Repeated fixtures, requests and seeds are grouped within physical scenes for inference. The separate command-duration follow-up contains 420 native attempts on the same configurations and seeds, with each command held for two environment steps.

File	Contents
`live_validation.json`	Complete live-study export, measurements, clustered analyses and raw-file checksums
`controller_timing.json`	Complete matched-time follow-up export and clustered comparisons
`live_reproduction.zip`	Both October studies: evaluator, analysis, tests, protocols, selections, traces, initial arrays, physics parameters, deduplicated initial images and excluded implementation checks
`round2_analysis.json`	Secondary request-swap contrasts, coverage/choice decomposition, benchmark-only sensitivity and fixture-level donor audit; unchanged rollout data
`round2_reproduction.zip`	Small analysis supplement: code, specification, tests and two manuscript table-panel generators
`confirmation.json`, `extension.json`	Unchanged September exports, with 320 and 60 scheduled attempts
`reproduction.zip`	September reproduction code, inputs, traces and separate development records
`SHA256SUMS.json`	Checksums of the files at this repository revision

Reproduce the live analysis

Download live_validation.json, controller_timing.json, and live_reproduction.zip from one fixed revision into the same directory. Extract the archive there and verify every file against its internal SHA256SUMS.json. Python 3.10 or later with NumPy, SciPy and pytest is sufficient; analysis does not require a GPU. The original analysis used Python 3.10.12, NumPy 1.25.0 and SciPy 1.14.0. Use these versions for byte-identical exports; other SciPy versions can differ in the last floating-point digits of descriptive beta-quantile intervals.

From the extracted directory:

cd code
export PYTHONPATH="$PWD/src:$PWD"
python -m pytest tests -q
python -m experiments.revision.materialize_recorded_images --root ..
python -m experiments.revision.analyze_live_validation \
  --root ../live --selection ../live/selection_v2.json --out ../checked_live.json
python -m experiments.revision.analyze_controller_timing \
  --root ../timing --selection ../timing/selection.json \
  --baseline-export ../live_validation.json --baseline-root ../live \
  --out ../checked_timing.json

Compare the checked files with the corresponding downloaded exports. Initial PNG files are stored once per content hash; the materialization command restores their episode paths before raw-file verification, without changing their bytes. The main analyzer requires all 930 result records per policy and checks the scheduled configurations, checkpoint/source hashes, consecutive traces, actual initial arrays, online stopping, reporting horizons and hold measurements. No failed trial is replaced. Goal attainment and retention during ten stationary-controller steps are separate outcomes. Blank and shuffled rollouts are each labelled against both candidate goals but count as one physical trial. Their pooled conditional target fraction is balanced by this two-label scoring; it is not an independent test of target preference.

To regenerate the nine October result tables, hash macros and native timing figure, install Matplotlib and run from code/:

python paper/reproduce/live_assets.py --live-export ../live_validation.json \
  --timing-export ../controller_timing.json --output ../manuscript_assets

Reproduce the secondary analysis

Download round2_analysis.json and extract round2_reproduction.zip into the same directory as the original archive. The supplement adds files without replacing the original evaluator or primary analyzer. From code/:

python -m experiments.revision.round2_analysis --export ../live_validation.json \
  --out ../checked_round2.json
ROUND2_PRIMARY_EXPORT=../live_validation.json ROUND2_SECONDARY_EXPORT=../round2_analysis.json \
  python -m pytest tests/test_round2_analysis.py -q
python -m paper.reproduce.round2_assets --export ../round2_analysis.json \
  --output ../manuscript_assets

Compare checked_round2.json with round2_analysis.json. These reviewer-requested secondary estimates reuse the original stratified scene bootstrap. They do not add rollouts or replace the four original primary tests. The specification defines the swap contrast, additive components, donor-overlap rule and benchmark-only restriction.

Rerun the live evaluation

Use separate simulator and policy environments, with versions and lock files under environments/. Install the pinned upstream code and assets:

VLABench: cf588fe60c0c7282174fe979f5913170cfe69017.
SmolVLA checkpoint: 4fd586e12dc14b04d9d606ddbb77448df4f0ff29.
pi0 checkpoint: 1ad73753a74d5cd97e67856664350f3f0baa21dc.
VLABench OpenPI implementation: 4483d1da6332da44115fe530e4e6fdd89bd57b13.

From code/, create fresh copies with local asset and scene paths:

python -m experiments.revision.prepare_reproduction --inputs ../inputs \
  --vlabench-root /path/to/VLABench/VLABench --out ../relocated
python -m experiments.revision.prepare_live_reproduction \
  --selection ../live/selection_v2.json --relocated-inputs ../relocated \
  --vlabench-root /path/to/VLABench/VLABench --out ../selection_local.json

Set PROJECT_STORAGE_ROOT to the absolute relocated/ path and VLABENCH_ROOT to the installed benchmark package. Keep PYTHONPATH set to code/src:code using absolute paths. Set MUJOCO_GL=egl, OMP_NUM_THREADS=4 and OPENBLAS_NUM_THREADS=4. Use one renderer at a time, with inference on separate GPUs. The supplied resource guard permits physical GPUs 0 through 4.

Start the services in their policy environments:

python -m experiments.e2a.smolvla_server --gpu 1 --port 5581
python -m experiments.revision.pi0_server --gpu 3 --port 5583 \
  --repo /path/to/openpi --checkpoint /path/to/pi0-checkpoint

Run the following commands serially in the simulator environment, with CUDA_VISIBLE_DEVICES=0, from code/:

python -m experiments.revision.live_validation --out ../rerun/smolvla \
  --selection ../selection_local.json --policy smolvla --port 5581 --chunk 50
python -m experiments.revision.live_validation --out ../rerun/pi0 \
  --selection ../selection_local.json --policy pi0 --port 5583 --chunk 5

Use fresh output directories; an existing manifest cannot be mixed with new settings. Seeds and jobs remain unchanged by path relocation. Native goals and both paired goals are checked online at each action. The 200-action common diagnostic limit is not every family's published task limit; the 100-action snapshot also reports the playing-card limit. A new simulator run need not reproduce complete trajectories bitwise. Keep its measurements separate from the downloaded frozen records.

For the command-duration follow-up, relocate the fixed native selection and run both policies serially, in separate fresh outputs, with the same service settings:

python -m experiments.revision.prepare_timing_reproduction \
  --selection ../timing/selection.json --vlabench-root /path/to/VLABench/VLABench \
  --out ../timing_selection_local.json
python -m experiments.revision.controller_timing --out ../timing_rerun/smolvla \
  --selection ../timing_selection_local.json --policy smolvla --port 5581 --chunk 50
python -m experiments.revision.controller_timing --out ../timing_rerun/pi0 \
  --selection ../timing_selection_local.json --policy pi0 --port 5583 --chunk 5

Physics and environment stepping remain at 1000 Hz and 10 Hz; commands are issued at 5 Hz. The common limit is 200 environment steps (20 s), at most 100 commands. Unchanged chunk counts double physical replanning intervals. This is a command-hold comparison, not an isolated frequency intervention. The timing analysis above verifies the downloaded original matched pairs; newly generated trajectories are separate measurements, not substitutes for these frozen records.

excluded/ preserves the implementation smoke checks and the interrupted launch's result-less initialization. These files are not additional study attempts. Completed main-study results are retained unchanged, including numerical failures.

The September exports and their original instructions remain available at the September revision.