---
license: mit
Modalities:
- image
- json
task_categories:
- question-answering
- summarization
language:
- en
tags:
- multimodal
- safety-critical scenario in AD
- Consistency-Centric Evaluation
- benchmarking
size_categories:
- n<1K
---

## RADIUS-Drive（及 RADIUS-Data）

本仓库在 Hugging Face 上发布 RADIUS-Drive（风险感知一致性诊断评测基准）。  

This Hugging Face repo hosts RADIUS-Drive, a risk-aware diagnostic benchmark for safety-critical autonomous driving.

- GitHub: https://github.com/pupprtseven/RADIUS_Drive_Bench
- Paper: 

## Benchmark Overview

RADIUS-Drive 关注 Pseudo-Correctness（伪正确性）：
模型可能给出表面正确的驾驶决策，但未能识别、定位或验证潜在风险机制。

RADIUS-Drive targets Pseudo-Correctness, where models produce correct-looking driving decisions while failing to trigger, localize, or validate underlying risk factors.

为此，RADIUS-Drive 提出 SAR（Safety → Awareness → Reasoning）诊断协议，
通过跨阶段一致性指标，系统性评估模型在安全关键决策中的 风险感知与推理一致性。

To this end, RADIUS-Drive introduces the SAR (Safety → Awareness → Reasoning) diagnostic protocol, enabling consistency-centric evaluation of risk perception and decision reasoning.

## 任务 / Tasks
### SAF（Safety）

检测模型是否能够识别并响应安全关键风险信号。  

Detect whether the model correctly triggers safety awareness under risk-critical conditions.

### AWR（Awareness）

评估模型是否能够定位并解释风险来源，而非仅给出结果。  

Assess whether the model localizes and explains the source of risk, beyond outcome correctness.

### REA（Reasoning）

验证模型的决策是否基于与风险一致的因果推理过程。  

Verify whether decisions are supported by risk-consistent causal reasoning.

### X-CONS（Cross-Phase Consistency）

检查 Safety、Awareness 与 Reasoning 之间是否存在一致性违背（Pseudo-Correctness）。  

Evaluate cross-phase consistency to diagnose pseudo-correct behavior.

## 关键约定 / Key Conventions

所有样本均配备 可审计的 ground truth，支持reference-based injection 与 reference-free generation 两种场景构造方式。  

All samples provide auditable ground truth, supporting both reference-based injection and reference-free generation.

## 评测结果应同时报告：

阶段内性能（per-phase）

跨阶段一致性指标（e.g., CF-CDA, Guess）

Results should report both per-phase scores and cross-phase consistency metrics.
## Dataset Contents
Each instance **X** is released as an image plus JSON sidecars:

- `dataX.png`: the rendered driving scene.
- `dataX.json`: taxonomy and post-decision supervision, including:
  - classification (Level-3)
  - `lt_ele` (dominant element)
  - `acc_factors`
  - `post_dec` (reference level and/or plan text)
- `dataX_aligned.json`: simulator-ready coarse state for Phase-1 Safety (map/relations/kinematics abstraction).
- `dataX_gt.json`: pre-decision action tags for Phase-1 scoring (per-action collision/hazard/safe, and optional best action).

## Design Principles
The schema is intentionally minimal:

- `dataX_aligned.json` is only as complex as needed for deterministic Safety rollout.
- `dataX.json` carries the Phase-2/3 labels required for SAR diagnosis and cross-phase consistency metrics.

## Directory Structure
```
RADIUS_DataSet550/
├── json/                # JSON sidecars for each instance
├── pic/                 # Rendered scene images
├── aligned_dataset.py   # Utility script for alignment
├── example.json
├── example_plot.png
└── README.md
```

## Notes
- File naming follows the `dataX.*` convention across image and JSON sidecars.
- For benchmark usage, refer to the `RADIUS_benchmark` module documentation.

## License
MIT License