---
language:
- en
license: cc-by-nc-sa-4.0
pretty_name: EgoPathBench
size_categories:
- 10K<n<100K
tags:
- embodied-ai
- spatial-reasoning
- visual-question-answering
- navigation
---

# EgoPathBench

Yang Zhao, Zhuo Chen, and Xubo Yang, Shanghai Jiao Tong University.

Release upload complete. All 36 archives were verified against SHA-256 checksums.
See `UPLOAD_COMPLETE.json` and `archive_manifest.json`.

31,852 train, 1,345 validation, and 1,111 benchmark questions across point and
embodied traversability, point and embodied paths, and intent paths.

## Loading

Download the full snapshot with `hf download runder1/EgoPathBench --repo-type dataset --local-dir downloads`.
Verify and extract the downloaded archives using `python scripts/unpack_dataset.py --source downloads --output dataset` from the code repository (Python 3.12+). Downloads require roughly 34 GB, plus another 34 GB for extraction. The 32 asset tar shards extract into one shared `assets/` tree.

This is a file-based release containing JSONL records and referenced assets; it is
not an automatically inferred `load_dataset` configuration.

Read `release/{split}/vqa/vqa_next_{task}.jsonl` for task questions.
Each row contains a question ID, prompts, a relative image path, visible-waypoint
metadata, and ground truth. Ground truth and evaluator sidecars are for evaluation
or training supervision only; never pass them to a zero-shot model.

**For benchmark scoring use `benchmark_evalfix/benchmark/vqa/` and its matching
`gt/` and `sidecars/`, rather than the unpatched benchmark labels.**
Run evaluation from this dataset root so asset paths resolve.

Original scene meshes, source-pool archives, provider credentials, and experiment
logs are not included. Historical construction provenance may contain original
machine paths; executable image/waypoint/evaluation references are local to this release.

The benchmark reference-route integrity check passed on all 819 route questions
after packaging. This validates packaging, not model performance.

## Training resource

`training/` contains 31,852 direct-answer and 31,852 spatial CoT training conversations, with portable image paths and the original export report.

## License and sources

CC BY-NC-SA 4.0, with applicable upstream terms retained. Scene geometry derives
through InternScenes from 3RScan, ScanNet, ARKitScenes, and Matterport3D.
See `THIRD_PARTY_NOTICES.md`. Original scene assets must be obtained upstream.

## Code and paper

[Code](https://github.com/EgoPathBench/EgoPathBench)

[Project page](https://egopathbench.github.io/EgoPathBench/)

[Paper](https://arxiv.org/abs/2609.16610) · arXiv:2609.16610

## Citation

```bibtex
@misc{zhao2026egopathbench,
  title={EgoPathBench: Evaluating Zero-Shot Egocentric Waypoint Decision-Making in Vision-Language Models},
  author={Yang Zhao and Zhuo Chen and Xubo Yang},
  year={2026},
  eprint={2609.16610},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.16610}
}
```


Contact: runder1103@sjtu.edu.cn; yangxubo@sjtu.edu.cn.