Datasets:

ReasonCore
/

open-spatial-reasoning

license: cc-by-4.0
task_categories:
  - multiple-choice
language:
  - en
tags:
  - autonomous-driving
  - spatial-reasoning
  - 3d-reasoning
  - vqa
pretty_name: Driving 3D Spatial Reasoning
size_categories:
  - n<1K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: image
      dtype: image
    - name: category
      dtype: string
    - name: question
      dtype: string
    - name: choices
      sequence: string
    - name: answer
      dtype: string

Open Spatial Reasoning

A multiple-choice dataset of spatial reasoning questions and answers for evaluating 3D spatial reasoning from single driving images. Each image contains numbered bounding boxes referencing objects in the scene, and each question probes a model's ability to reconstruct the real 3D scene rather than rely on flat-image shortcuts (e.g. "lower in the frame = closer", "bigger box = nearer").

Dataset Description

Frontier vision-language models often answer these questions correctly by luck while reasoning incorrectly, leaning on pixel-layout heuristics that break down on elevated roads, slopes, curves, and intersections. This dataset is designed to surface that failure mode by requiring metric 3D reasoning about distance, lateral position, ordering, and heading.

Each sample pairs a driving-scene image with a question, four answer choices, and the correct answer letter.

The images were collected by autonomous vehicles operated by PlusAI.

Data Fields

Field	Type	Description
`id`	`string`	Unique question identifier (e.g. `q0002`)
`image`	`image`	The driving image with numbered bounding boxes
`category`	`string`	The reasoning task type (see categories below)
`question`	`string`	The full question, including the reasoning protocol
`choices`	`list[string]`	Four answer options, prefixed `A.`–`D.`
`answer`	`string`	The correct answer letter (`A`, `B`, `C`, or `D`)

Question Categories

The dataset spans several spatial-reasoning task types, including:

Category	What it tests
`identify_distance_long`	Estimate the absolute distance to an object (binned 0–20m / 20–50m / 50–80m / 80m+)
`relative_distance_long`	Estimate the 3D separation between two objects
`pick_closer`	Decide which of two objects is closer to the ego vehicle
`identify_rightmost`	Identify the object furthest to the right in true 3D space
`order_leftmost`	Order several objects left-to-right in 3D space
`identify_position`	Classify an object's position relative to ego (e.g. ahead-left, behind-right)
`identify_heading`	Determine an object's heading using clock directions (12 = forward, 3 = right)

Authors

Anurag Ganguli, Anshuman Lall, Abhishek Bhatia, Xiangyu Gao, Joe Yuan, Satish Vutukuru, Geoff Wolfe

Citation

If you use this dataset, please cite it:

@misc{driving_3d_spatial_reasoning,
  title  = {Open Spatial Reasoning},
  author = {Anurag Ganguli, Anshuman Lall, Abhishek Bhatia, Xiangyu Gao, Joe Yuan, Satish Vutukuru, Geoff Wolfe},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/reasoncore/open-spatial-reasoning}}
}