Datasets:

nur-dev
/

robot-execution-validation

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

Robot Execution Validation

VLABench execution records for released SmolVLA and Pi0 policies: 320 primary confirmation attempts, a prospectively specified 60-attempt family-matched native validation extension, and 37 separately reported development attempts. Each native trajectory is evaluated at 60 and 200 actions. Paired instructions use the same initial state and requested-versus-alternative goal predicates.

File	Contents
`confirmation.json`	Frozen primary results, per-trial measurements, summaries, and raw-file hashes
`extension.json`	Frozen validation-extension results and the prospective configuration selection
`reproduction.zip`	Evaluation/analysis code, tests, protocols, selected inputs, complete action traces, initial states, scene images, and environment versions
`SHA256SUMS.json`	File checksums

Reproduce the analysis

Download the files from one fixed repository revision and extract reproduction.zip. The archive has its own file-level SHA256SUMS.json. Python 3.10 or later with NumPy, SciPy, Matplotlib, and pytest is sufficient for analysis; no GPU is required. Use the recorded package versions for exact regeneration.

From the extracted directory:

cp /path/to/download/confirmation.json code/paper/reproduce/revision_results.json
cp /path/to/download/extension.json code/paper/reproduce/validation_extension_results.json
cd code
export PYTHONPATH="$PWD/src:$PWD"
python -m pytest tests -q
python -m experiments.revision.freeze --root ../confirmation --out ../confirmation_check.json
python -m experiments.revision.freeze_extension --root ../extension --out ../extension_check.json
python -m paper.reproduce.revision_assets --out ../latex_assets
python -m paper.reproduce.extension_assets --out ../latex_assets

The checked exports reproduce the downloaded JSON files. The generated assets include the original four-family native results, paired outcomes, progress, development results, family-matched extension and synthesis, and stopping-rule sensitivity. Technical errors remain distinct from observed goal outcomes. The paired first-goal analysis is post hoc; the original native-stop outcomes are retained alongside it.

Extension validation compares the frozen selected configurations with the recorded inputs, undoing only VLABench's automatic expansion of relative XML asset paths. No raw record or prospective selection is changed. The extension retains three Pi0 numerical failures.

Rerun execution

Use the package inventories in environments/versions.json and install these pinned upstream sources and their assets in separate simulator and policy environments:

VLABench, revision cf588fe60c0c7282174fe979f5913170cfe69017.
SmolVLA checkpoint, revision 4fd586e12dc14b04d9d606ddbb77448df4f0ff29.
Pi0 checkpoint, revision 1ad73753a74d5cd97e67856664350f3f0baa21dc.
Author OpenPI implementation, revision 4483d1da6332da44115fe530e4e6fdd89bd57b13; use the included openpi_uv.lock.

Initialize a local Git checkout of code/ before execution so the evaluator can record the reproduction's source revision. From code/, relocate archived asset references into a fresh directory:

python -m experiments.revision.prepare_reproduction --inputs ../inputs \
  --vlabench-root /path/to/VLABench/VLABench --out ../relocated

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

Start the services in their respective 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 evaluator in the simulator environment with CUDA_VISIBLE_DEVICES=0. For the extension, use --tasks select_book,select_mahjong,select_poker and the relocated published track. For the original native control, use --tasks select_fruit,select_toy,select_chemistry_tube,select_drink.

python -m experiments.revision.vlabench_control --out /path/to/fresh/run \
  --tasks select_book,select_mahjong,select_poker \
  --track-config ../relocated/track_1_in_distribution.json --episodes-per-task 10 \
  --policy-seed 17 --max-substeps 1 --max-steps 200 --port 5581 --chunk 50

Use --port 5583 --chunk 5 for Pi0. For paired instructions, replace --tasks and --track-config with --paired-spec ../relocated/final_study_command_spec.json. Use fresh output directories and compare actual initial-state and observation hashes. The common 200-action diagnostic budget is not every task's official leaderboard budget.

The separate synthetic-speech records are available in multilingual-robot-grounding-causal-audit.