Hugging Face's logo

Datasets:
canyuchen
/
embodied-safety-exam-renders

Size:
n<1K
Tags:
robotics
embodied-ai
safety
mujoco
blender
Dataset card Files
xet
 Community
1
Dataset Viewer
Duplicate
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
suite: string
task: string
source: string
model: string
board_run: string
board_trial_id: string
board_reward: double
pass: bool
outcome: string
judgment_class: string
renderer: string
rendering: string
width: int64
height: int64
fps: int64
speed: double
image: string
rcb_commit: string
note: string
former_task: string
former_image: string
rcb_task: string
former_rcb_task: string
success_check: bool
to
{'suite': Value('string'), 'task': Value('string'), 'source': Value('string'), 'renderer': Value('string'), 'rendering': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'fps': Value('int64'), 'speed': Value('float64'), 'success_check': Value('bool'), 'image': Value('string'), 'rcb_task': Value('string'), 'rcb_commit': Value('string'), 'outcome': Value('string'), 'note': Value('string'), 'former_rcb_task': Value('string'), 'former_image': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              suite: string
              task: string
              source: string
              model: string
              board_run: string
              board_trial_id: string
              board_reward: double
              pass: bool
              outcome: string
              judgment_class: string
              renderer: string
              rendering: string
              width: int64
              height: int64
              fps: int64
              speed: double
              image: string
              rcb_commit: string
              note: string
              former_task: string
              former_image: string
              rcb_task: string
              former_rcb_task: string
              success_check: bool
              to
              {'suite': Value('string'), 'task': Value('string'), 'source': Value('string'), 'renderer': Value('string'), 'rendering': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'fps': Value('int64'), 'speed': Value('float64'), 'success_check': Value('bool'), 'image': Value('string'), 'rcb_task': Value('string'), 'rcb_commit': Value('string'), 'outcome': Value('string'), 'note': Value('string'), 'former_rcb_task': Value('string'), 'former_image': Value('string')}
              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.

Embodied Safety Exam 0.5.4: cinematic renders

Path-traced videos of the 84 core tasks of the Embodied Safety Exam (ESE) 0.5.4. The benchmark was called RoboSafeBench when these tasks were merged into robot_coding_bench (RCB, now embodied-first-exam) in PR #296 (RCB commit baf9e716, task prefix rsb-). Each video is a faithful replay of a recorded run, re-lit and re-framed for viewing.

This dataset was canyuchen/robosafebench-renders (that old address no longer redirects). On the rename the folders went from oracle/rsb-… to oracle/ese-… and meta.json got the new names.

2026-10-30: all 84 oracle videos re-rendered. A frame-by-frame audit of the first renders found single-frame material swaps (Blender persistent data), z-fighting flicker on coincident faces, black strips from room walls that were hidden from the camera but still blocked light, and grey glass, dark liquids and hard-edged smoke. All of that is fixed in oracle/; the replayed motion is the same, some speeds changed (see speed in meta.json), and each note ends with a re-render sentence.

2026-11-03: all 228 trial videos re-rendered with the same fixes. The replayed motion and the speeds are the same as in the 2026-10-19 renders, and each note ends with a re-render sentence. The render audit of these videos covers sampled frames, a flicker check and an automated visual review of each video, with disputed findings checked against the replay data. Some oddities remain because they come from the assets or the recorded run, not the render: liquids stay as authored and can look opaque, a robot link can pass through another part where the replay has it there, and a few props look unfinished.

Contents

  • oracle/<task>/ for all 84 tasks: video.mp4 (2560x1440, 30 fps, H.264 CRF 16), poster.png, meta.json. The task name is the ESE task id, e.g. ese-c16-i00-standard or ese-c02-i00-standard-realtime (the RCB #296 id with rsb- replaced by ese-). 76 of the 84 are in embodied-first-exam today (the 8 that GPT-6 Luna solved were left out).
  • check_media_exam_suite.txt: the output of the media-standard v2 checker on the 79-task exam suite built from these videos (re-run on the 2026-10-30 renders: 0 failing, 0 warnings).
  • trials/<model>/<task>/ for 228 board trials of three models (astra = GPT-6 Astra via Codex, opus55 = Claude Opus 5.5 via Claude Code, gemini4 = Gemini 4 via Antigravity CLI; 76 tasks each), same files and render rules. These are model runs, passed or not: pass, outcome, judgment_class and board_reward in meta.json are the live verdict, which the in-image replay reproduced (same end-state hash, same verdict) for every one. former_task keeps the RoboSafeBench name.

How the videos were made

  1. Replay. The journal of the task's Harbor oracle trial is re-executed inside the task image (ghcr.io/mll-lab-nu/eai-robosafe:0.5.4 at the time, the image now tagged efe-private:ese-0.5.4) with the image's own simulator package (MuJoCo 3.8.1), as the benchmark's sim service does. For every task the end-state hash equals the live record's and the verdict equals the live one (success_check: true; outcome safe_done or correct_stop in meta.json).
  2. Poses. The body poses of every 20 ms control tick are written into Blender as keyframes. Nothing is simulated in Blender; geometry, textures and visual events (colour changes, smoke, liquids appearing) come from the MuJoCo model of the task (asset profile realism_v3).
  3. Render. Blender 4.5.14 LTS, Cycles path tracer, 64 samples adaptive (threshold 0.015), OpenImageDenoise with albedo and normal passes, AgX Medium High Contrast, motion blur 0.25 frame. Studio sweep with a warm key, top light, cool rim and weak fill; 3/4 camera with a slow 10 degree orbit. The 15 room tasks use a chase camera; walls between camera and robot are left out of the render (whole walls, cutaway view). meta.json records the device (OptiX on an NVIDIA L40S or GeForce RTX 4090).
  4. Timing. Constant speed per task, 1x to 8x (speed in meta.json); first frame held 0.6 s.

meta.json fields

suite (ese), task (base id), source (oracle), renderer, rendering (path-tracing), width, height, fps, speed, success_check, image, former_image, rcb_task, former_rcb_task (the rsb- id of RCB #296), rcb_commit, outcome, note (how the replay was checked).

Related

Raw board results of the model trials on these tasks: https://huggingface.co/datasets/canyuchen/embodied-safety-exam-board-results
Downloads last month
189
Total file size:
5.34 GB