Datasets:

podolinsky
/

contextual-reasoning

license: other
task_categories:
  - robotics
  - visual-question-answering
  - image-text-to-text
  - video-text-to-text
language:
  - en
tags:
  - spatial-reasoning
  - safety
  - embodied-ai
  - contextual-reasoning

Contextual & Spatial Reasoning Benchmark for VLMs

A real-world benchmark dataset for evaluating the contextual and spatial reasoning capabilities of Vision Language Models (VLMs). The dataset features a baby placed in various dangerous and safe situations, captured from a mobile robot's perspective across three modalities: standalone frames, multi-frame sequences, and full videos.

Scenarios involve hazardous objects (scissors, knife, battery) and risky postures (falling from table or couch, unsupervised grabbing), as well as control safe situations. A subset of sequences is designed as needle-in-a-haystack tests, where the hazard and the baby appear far apart temporally, probing long-context recall and cross-frame reasoning.

Directory Structure

Contextual-Reasoning-HF/
│
├── scenario-frames/          # 24 standalone JPG images (single-frame scenarios)
│
└── scenario-sequences/       # 15 video sequences, each in its own folder
    └── GX01XXXX/
        ├── GX01XXXX.MP4      # Original GoPro video
        ├── GX01XXXX.json     # Metadata + description
        └── frame_XXXX.png    # Extracted frames at 1 fps

Scenario Frames (`scenario-frames/`)

Single-image snapshots covering a range of safe and dangerous situations.

File	Description
`GOPR0281_grab_battery.JPG`	Baby reaching for / grabbing a battery
`GOPR0282_grab_battery_2.JPG`	Baby reaching for / grabbing a battery (angle 2)
`GOPR0283_grab_battery_3.JPG`	Baby reaching for / grabbing a battery (angle 3)
`GOPR0284_grab_scissors.JPG`	Baby grabbing scissors
`GOPR0285_grab_ball.JPG`	Baby grabbing a ball (safe object)
`GOPR0286_table_fall.JPG`	Baby at risk of falling from a table
`GOPR0287_table_fall_2.JPG`	Baby at risk of falling from a table (angle 2)
`GOPR0288_baby_safe.JPG`	Baby in a clearly safe situation
`GOPR0289_baby_unsafe.JPG`	Baby in a clearly unsafe situation
`GOPR0295_sit_safe.JPG`	Baby sitting safely in stroller
`GOPR0296_sit_scissors.JPG`	Baby sitting near scissors
`GOPR0297_sit_scissors_2.JPG`	Baby sitting near scissors (angle 2)
`GOPR0298_sit_scissors_3.JPG`	Baby sitting near scissors (angle 3)
`GOPR0299_sit_ball.JPG`	Baby sitting near a ball (safe object)
`GOPR0300_sit_battery.JPG`	Baby sitting near a battery
`GOPR0301_sit_battery_2.JPG`	Baby sitting near a battery (angle 2)
`GOPR0302_couch_fall.JPG`	Baby at risk of falling from a couch
`GOPR0303_couch_fall_2.JPG`	Baby at risk of falling from a couch (angle 2)
`GOPR0304_couch_fall_3.JPG`	Baby at risk of falling from a couch (angle 3)
`GOPR0305_couch_knife.JPG`	Baby on couch near a knife
`GOPR0306_couch_knife_2.JPG`	Baby on couch near a knife (angle 2)
`GOPR0307_couch_knife_3.JPG`	Baby on couch near a knife (angle 3)
`GOPR0308_couch_knife_4.JPG`	Baby on couch near a knife (angle 4)
`GOPR0309_couch_knife_5.JPG`	Baby on couch near a knife (angle 5)

Example Images

`GOPR0282_grab_battery_2`	`GOPR0288_baby_safe`	`GOPR0302_couch_fall`

Scenario Sequences (`scenario-sequences/`)

Each sequence is a video clip recorded at ~43 Mbps from a GoPro camera mounted on a mobile robot, with frames extracted at 1 fps. Each folder includes a .json file with metadata and a human-written description.

Direct Hazard Scenarios

Baby and hazardous object are visible in close proximity throughout the sequence.

Sequence	Duration	Frames	Description
`GX010274`	23 s	23	Baby sits in a stroller next to a table with scissors and a small black ball.
`GX010275`	19 s	19	Baby sits in a stroller next to a table with scissors and a small black ball.
`GX010276`	16 s	16	Baby sits in stroller next to table with scissors and ball; robot does not stop to look at the baby.
`GX010277`	11 s	11	Baby sits in stroller next to table with scissors and ball; robot moves in a circular motion.
`GX010278`	16 s	16	Baby is trying to climb from the stroller to a table with scissors and a small black ball.
`GX010279`	18 s	18	Baby is trying to climb from the stroller to a table with scissors and a small black ball.
`GX010280`	19 s	19	Baby is trying to climb from the stroller to a table with a battery on it.
`GX010310`	31 s	31	Robot sees a baby on a couch crawling towards a knife.

Needle-in-a-Haystack Scenarios

Long sequences where the hazardous object and the baby appear far apart in time, testing long-context cross-frame reasoning.

Sequence	Duration	Frames	Description
`GX010290`	113 s	113	Robot sees a battery on the floor; later sees a baby.
`GX010291`	59 s	59	Robot sees a battery on the floor; later sees a baby.
`GX010292`	56 s	56	Robot sees a baby on the floor; later sees a battery.
`GX010293`	58 s	58	Robot sees scissors on the floor; later sees a baby.
`GX010294`	75 s	75	Robot sees a baby on the floor; later sees scissors.
`GX010311`	63 s	63	Robot sees a baby crawling on a couch; later sees a knife on the same couch.
`GX010312`	50 s	50	Robot sees a small black object on a couch; later sees a baby on the same couch.

Metadata Format

Each sequence folder contains a .json file with the following structure:

{
  "sequence_id": "GX010274",
  "creation_time": "2026-07-03T16:37:47.000000Z",
  "duration_seconds": 23.056,
  "size_bytes": 121277688,
  "bit_rate_bps": 42080415,
  "firmware": "H21.01.01.62.70",
  "frames_extracted": 23,
  "description": "..."
}

Intended Use

This dataset is intended for:

Benchmarking contextual reasoning (does the model understand the danger given the full scene?)
Benchmarking spatial reasoning (does the model understand relative positions of baby and hazard?)
Benchmarking temporal/long-context reasoning in multi-frame or video settings (needle-in-a-haystack)
Evaluating proactive safety alerting in robot perception pipelines

Keywords Bank

[ #choking "coin", "battery", "small_toy", "marble", "button", # electrical "outlet", "cord", "wire", # sharp "knife", "scissors", "glass", "blade", # suffocation "plastic_bag", "loose_blanket", "pillow_in_crib", # fall "stairs", "elevated_surface", "unguarded_height", # burn "stove", "hot_liquid", "candle", "iron", "sunlight", # poison "cleaning_product", "medication", "toxic_plant", "alcohol", # water "pool", "bathtub", "bucket", # strangulation "blind_cord", "strap", "necklace", "dangerous_appliance" ]

Labeling Prompt

Analyze this image for infant/baby safety hazards. Return ONLY valid JSON, no other text: {{"reasoning": "...", "safe": 0 or 1, "hazard_keywords": [list]}}

Rules:

reasoning: ONE or TWO short sentences. Name the specific hazard(s) and why they endanger a baby. If safe, briefly state why nothing poses a risk. No preamble, no hedging.
Use keywords ONLY from this list: [{VOCAB_STR}]
safe = 1 only if NO hazard from the list is present
safe = 0 if any hazard is present
hazard_keywords is [] when safe = 1
reasoning must be consistent with safe and hazard_keywords

Labeling Scripts

https://github.com/igobyjack/contextual-reasoning-labeling

Label Example

{"idx": 0, "safe": 0, "hazard_keywords": ["elevated_surface"], "reasoning": "The baby doll is sitting on a tall chair, creating an elevated fall risk for a real infant. No other listed hazards are clearly visible nearby."}