Hugging Face's logo

Datasets:
nvidia
/
Arena-DROID-Camera-Sensitivity-Workflow-Sample

Tasks:
Robotics
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
robotics
simulation
policy-evaluation
sensitivity-analysis
isaaclab-arena
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
Arena-DROID-Camera-Sensitivity-Workflow-Sample / README.md
ClemensCV's picture
ClemensCV
Add camera sensitivity dataset card (#1)
1667456
preview code
|
Raw
History Blame Contribute Delete
9.02 kB
metadata 
pretty_name: Arena DROID Camera Sensitivity Workflow Sample
task_categories:
  - robotics
tags:
  - robotics
  - simulation
  - policy-evaluation
  - sensitivity-analysis
  - isaaclab-arena
language:
  - en
size_categories:
  - n<1K

Arena DROID Camera Sensitivity Workflow Sample

Dataset Description

Arena-DROID-Camera-Sensitivity-Workflow-Sample is a compact set of episode-level results generated by an Isaac Lab-Arena simulation experiment. It lets users run the documented camera sensitivity analysis without first executing the policy-evaluation sweep.

The experiment evaluates an OpenPI pi05 policy on a DROID Rubik's-cube pick-and-place task while independently varying the wrist-camera position on three axes. Each JSONL record contains the sampled camera offset, task instruction, episode outcome and length, timestamp, objective progress, and predicate events. The dataset does not contain robot trajectories, actions, observations, images, video, audio, or model weights.

Dataset version Records Successful Unsuccessful Data file
arena_v0.2_lab_v3.0 10 7 3 episode_results_rebuild0.jsonl

This is a workflow sample. Its size is sufficient to demonstrate the software path, but not to measure policy robustness or compare policy performance.

Dataset Owner

Dataset Creation Date

  • Simulation run: 2026-08-06
  • Hugging Face repository created: 2026-08-07

License/Terms of Use

No dataset-specific license is currently declared for this repository.

Intended Use

The dataset is intended to:

  • reproduce the documented Isaac Lab-Arena sensitivity-analysis tutorial without rerunning the evaluation;
  • exercise the episode-results reader and sensitivity-report generation path; and
  • illustrate the episode-result schema produced by a controlled environment-variation sweep.

The dataset is not intended for:

  • training or fine-tuning a robot policy;
  • benchmarking, ranking, or comparing policies;
  • drawing conclusions about policy robustness, safety, or real-world performance; or
  • validating a production or safety-critical robotics system.

Dataset Characterization

Data Collection Method

The records were generated automatically in simulation. No human teleoperation or manual annotation was used during this evaluation run or included in the released payload. This statement does not describe the separate training data used to develop the OpenPI policy.

The experiment used:

  • Isaac Lab-Arena's DROID absolute-joint-position embodiment;
  • an OpenPI pi05 policy through the Arena OpenPI adapter;
  • the instruction, "Pick up the Rubik's cube and place it in the bowl.";
  • a simulated Rubik's cube, bowl, maple table, and home-office environment;
  • 10 episodes in one simulation environment with seed 42; and
  • independent wrist-camera translation offsets sampled from -0.03 m to 0.03 m on each of three axes.

The dataset does not include the simulation assets, OpenPI software, or policy weights used during generation.

Labeling Method

Isaac Lab-Arena generated the success, progress, and event values automatically from the task's objective predicates. There are no human-provided labels.

Version and Provenance

  • Dataset revision: arena_v0.2_lab_v3.0
  • Dataset content commit: 6a879e4b09bc9911453dd8271081dfcf7d3aa9e1
  • Data-file SHA-256: 329a25484c2edc997d74e540b3bfe615a99390270c696c29effb73cc86497ea8
  • Isaac Lab-Arena introduction: pull request 1055

Dataset Format

The dataset contains one newline-delimited JSON file, episode_results_rebuild0.jsonl. Each line is one completed episode.

Field Type Description
job_name string Experiment run name.
env_id integer Simulation environment index.
episode_in_env integer Zero-based episode index within the environment.
seed integer Experiment seed recorded for the episode.
success boolean Automatically evaluated task-success outcome.
episode_length integer Episode length in environment steps.
language_instruction string Instruction supplied to the policy.
timestamp string Episode timestamp in ISO 8601-like format.
variations object Sampled environment-variation values.
progress object Objective scores, completion state, active predicates, and predicate events.

The variations object contains droid_abs_joint_pos.camera_extrinsics_wrist_camera, a three-element floating-point vector in metres. In the camera optical frame:

  • element 0 is horizontal displacement: negative is left and positive is right;
  • element 1 is vertical displacement: negative is up and positive is down; and
  • element 2 is depth displacement: negative is backward and positive is forward.

The progress.events array records the episode-local environment step, objective, predicate, and score change for each completed task event.

Hugging Face may present the records as an automatically inferred train split. That label is a viewer convention; these are evaluation results, not training data.

Dataset Quantification

  • JSONL records: 10
  • Data-file size: 12,101 bytes
  • Successful episodes: 7
  • Unsuccessful episodes: 3
  • Environment IDs: one (env_id = 0)
  • Episode indices: 0 through 9
  • Recorded run interval: 2026-08-06 10:43:07 through 10:57:21
  • Images, videos, and audio: none

The released JSONL parses as 10 complete JSON objects. The recorded camera offsets are within the configured [-0.03, 0.03] m bounds.

Personal Data and Third-Party Content

A review of episode_results_rebuild0.jsonl at revision arena_v0.2_lab_v3.0 found no names, email addresses, account identifiers, faces, voices, or real-world sensor captures. The file contains experiment identifiers, fixed task text, numeric results, and timestamps without timezone offsets.

The JSONL contains textual identifiers referring to DROID, OpenPI, and a Rubik's cube. It does not include or redistribute related assets, software, or model weights.

Limitations

  • Ten episodes are too few for reliable sensitivity, robustness, or performance conclusions.
  • All records come from one task, policy variant, seed, and simulation configuration.
  • The dataset contains episode summaries rather than the raw actions, observations, camera frames, or simulator state needed to reproduce individual rollouts.
  • Outcomes depend on the task predicates and thresholds used by the recorded workflow.
  • Simulation results do not establish real-world robot behavior or safety.
  • The timestamps do not include an explicit timezone offset.
  • The JSONL does not identify the exact OpenPI model or software commit used to generate the run.
  • The downloadable sample has a 0.7 success rate. Metrics shown for other runs in the workflow documentation do not describe this file.

Download and Use

Download the version used by the documented workflow:

hf download \
  nvidia/Arena-DROID-Camera-Sensitivity-Workflow-Sample \
  episode_results_rebuild0.jsonl \
  --repo-type dataset \
  --revision arena_v0.2_lab_v3.0

Follow the Isaac Lab-Arena sensitivity-analysis workflow to generate the example posterior report.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and has established policies and practices to support the development of AI applications. Users should evaluate whether this dataset and the surrounding workflow are appropriate for their use case and should not treat this small simulation sample as evidence of safe or effective deployment.

Please report security vulnerabilities or NVIDIA AI concerns through the NVIDIA Product Security reporting page.

References