README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- robotics
- video-classification
- visual-question-answering
language:
- en
tags:
- robotics
- egocentric-video
- video-understanding
- action-segmentation
- subtask-annotation
- benchmark
pretty_name: What's Going On Bench
---

# What's Going On Bench

Initial benchmark slice for evaluating VLM-generated task, subtask, and frame-level action annotations against RoboInter gold labels.

The main table is:

```text
whats_going_on_bench.parquet
```

It has one episode per row.

This first release contains 2,000 RoboInter LeRobot v2.1 episodes:

- 1,000 DROID episodes from `lerobot_droid_anno`
- 1,000 RH20T episodes from `lerobot_rh20t_anno`
- only episodes where `instruction_quality == "Primary"` and `subtask_quality == "Primary"`
- random chunk-balanced selection with seed `20260520`
- per-frame labels and subtask segments
- primary camera MP4 bytes embedded directly in each row

## Main Schema

Each parquet row is one episode:

```text
bench_id: string
source_dataset: string                  # droid | rh20t
lerobot_dataset: string
chunk_id: int32
episode_index: int32
episode_name: string
camera_view: string
fps: float32
num_frames: int32
duration_sec: float32
instruction: string
instruction_quality: string             # Primary | Secondary | empty
subtask_quality: string
primitive_skill_quality: string
subtasks: list<struct<
  start_frame: int32,
  end_frame: int32,
  start_sec: float32,
  end_sec: float32,
  subtask: string,
  primitive_skill: string
>>
frame_annotations: list<struct<
  frame_index: int32,
  time_sec: float32,
  subtask: string,
  primitive_skill: string
>>
primary_video: binary
primary_video_mime: string              # video/mp4
primary_video_sha256: string
upstream_repo: string
```

There are no external video paths needed for normal benchmark use.

## Selection

Rows are sampled from random RoboInter LeRobot chunks, then filtered to Primary/Primary labels:

```text
seed: 20260520
droid_chunks: [78, 107]
rh20t_chunks: [20, 28]
instruction_quality: Primary
subtask_quality: Primary
```

## Annotation Fields

The benchmark currently focuses on:

- `annotation.instruction_add`: episode/task instruction
- `annotation.time_clip`: subtask temporal segments in frame indices
- `annotation.substask`: current subtask label
- `annotation.primitive_skill`: optional low-level primitive label
- `Q_annotation.*`: Primary/Secondary quality flags

The source videos run at 10 FPS.

## Source And License

This dataset is derived from [InternRobotics/RoboInter-Data](https://huggingface.co/datasets/InternRobotics/RoboInter-Data), which is released under CC BY-NC-SA 4.0. Keep downstream usage consistent with the upstream license and attribution requirements.