README.md

metadata

license: mit
configs:
  - config_name: default
    data_files:
      - split: spatiotemporal_planning
        path: data/spatiotemporal_planning-*
      - split: embodied_grounding
        path: data/embodied_grounding-*
  - config_name: spatial_reasoning
    data_files:
      - split: spatial_reasoning
        path: data/spatial_reasoning-*
dataset_info:
  - config_name: default
    features:
      - name: question_type
        dtype: string
      - name: image
        dtype: image
      - name: question
        dtype: string
      - name: answer
        dtype: string
    splits:
      - name: spatiotemporal_planning
        num_bytes: 666694379.4
        num_examples: 49800
      - name: embodied_grounding
        num_bytes: 1870487844
        num_examples: 100000
    download_size: 2498651415
    dataset_size: 2537182223.4
  - config_name: spatial_reasoning
    features:
      - name: video
        dtype: string
      - name: question
        dtype: string
      - name: answer
        dtype: string
      - name: question_type
        dtype: string
    splits:
      - name: spatial_reasoning
        num_bytes: 36647782
        num_examples: 101000
    download_size: 6232491
    dataset_size: 36647782
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - image-text-to-text
tags:
  - hunyuan
  - vision-language-model
  - vision-language-action
  - embodied
  - Spatial-Intelligence

GEM: Generative Supervision Helps Embodied Intelligence

Ruowen Zhao¹, Bangguo Li¹, Zuyan Liu^1,2,†, Yinan Liang¹, Junliang Ye¹, Fangfu Liu¹,
Diankun Wu¹, Zhengyi Wang¹, Xumin Yu², Yongming Rao^2,✉, Han Hu², Jun Zhu^1,✉
†Project Lead.^✉Corresponding Author.
¹Tsinghua University, ²Tencent Hunyuan

Abstract

Embodied Vision-Language Models (VLMs) have demonstrated impressive performance and generalization in robotics, particularly within Vision-Language-Action frameworks. However, a significant gap remains between the high-level semantic focus of standard text-guided pre-training paradigms and the low-level spatial and physical knowledge critical for execution in embodied environments. In this paper, we introduce GEM, a Generative-supervised Embodied vision-language Model designed to bridge this divide. We propose integrating a depth map generation task directly into the VLM pre-training phase. By training this generative objective jointly with the main model, we observe substantial improvements in embodied intelligence, significantly enhancing both semantic understanding and physical operation capabilities. To support this paradigm, we curate and release GEM-4M, a comprehensive large-scale dataset featuring a mixture of grounding, reasoning, and planning data paired with high-quality depth supervision. Extensive experiments demonstrate that GEM achieves state-of-the-art results across diverse embodied benchmarks. Furthermore, our deployed action model, GEM-VLA, exhibits vastly superior task execution abilities in both simulation environments and real-world evaluations.

Dataset Structure

GEM-250K/
├── assets/
│   ├── GEM-demo.mp4
│   ├── overview.png
├── data/
│   ├── embodied_grounding-00000-of-00002.parquet
│   ├── embodied_grounding-00001-of-00002.parquet
│   ├── spatial_reasoning-00000-of-00001.parquet
│   └── spatiotemporal_planning-00000-of-00001.parquet
└── README.md

Citation

If you find this dataset helpful, please cite:

@article{zhao2026gem,
  title={GEM: Generative Supervision Helps Embodied Intelligence},
  author={Zhao, Ruowen and Li, Bangguo and Liu, Zuyan and Liang, Yinan and Ye, Junliang and Liu, Fangfu and Wu, Diankun and Wang, Zhengyi and Yu, Xumin and Rao, Yongming and others},
  journal={arXiv preprint arXiv:2605.28548},
  year={2026}
}