Figure 1: Overview of the RoboGene dataset.
**RoboGene** is a large-scale robotic task dataset generated by a **diversity-driven agentic framework**. Specifically designed for **Vision-Language-Action (VLA)** model pre-training, it addresses the critical bottlenecks of limited scene variety and insufficient physical grounding in existing datasets, significantly enhancing the generalization capabilities of humanoid robots in complex real-world tasks. ## Dataset Details ### Dataset Description RoboGene overcomes the limitations of traditional human teleoperation for data collection. By leveraging foundation models (e.g., GPT-4o) as generators within the **RoboGene Agent Loop**—which integrates retrieval, generation, and self-reflection—the framework ensures that every generated instruction is not only semantically diverse but also physically executable in structured environments. - **Curated by:** Beijing Innovation Center of Humanoid Robotics, Peking University, Beihang University, Beijing Institute of Technology. - **Shared by:** RoboGene Project Team - **Language(s) (NLP):** Chinese (Task Instructions) - **License:** Apache-2.0 ### Dataset Sources - **Demo:** (https://robogene-boost-vla.github.io/) - **Paper:** *RoboGene: Boosting VLA Pre-training via Diversity-Driven Agentic Framework for Real-World Task Generation* - **Framework Compatibility:** Native support for the [LeRobot](https://github.com/huggingface/lerobot) framework. ## Uses ### Direct Use This dataset serves as a **pre-training foundation** for VLA models: - **Massive-Scale Pre-training:** Enhances model robustness against novel objects, lighting variations, and background clutter. - **Zero/Few-shot Adaptation:** Experiments show that models pre-trained on RoboGene require only 50 human demonstrations to quickly adapt to complex downstream tasks. - **Dual-arm Coordination Research:** Contains high-quality trajectories for sophisticated dual-arm manipulation. ### Out-of-Scope Use - Deploying pre-trained policies directly to heavy industrial equipment without safety validation or simulation-based verification is strictly prohibited. ## Dataset Structure The dataset is organized by scenes with a balanced distribution. Each episode includes: - **Observations:** Multi-view RGB images (front-view, wrist-view) and proprioceptive states. - **Actions:** High-frequency control sequences for joints or end-effectors. - **Instructions:** Verified, high-quality natural language task descriptions. ### Task Highlights - **100+ Diverse Skills:** Ranging from basic pick-and-place to complex multi-stage assembly and tool usage. - **Rich Scenarios:** Comprehensive coverage of Kitchen, Laboratory, Medical, Office, Retail, Education, Industrial, and Daily Life environments. - **Complex Object Spectrum:** Includes high-reflectivity materials (e.g., glass), deformable/flexible objects (e.g., cloth, ropes), and articulated/hinged mechanisms (e.g., doors, drawers, cabinets). ## Dataset Creation ### Curation Rationale While human-curated data is physically valid, it often lacks structural diversity. Conversely, traditional LLM-generated tasks frequently fail to align with physical constraints. RoboGene bridges this gap by embedding physical grounding into an automated generation loop, resulting in pre-training performance that significantly outperforms both purely human-curated and baseline foundation-model datasets. ## Bias, Risks, and Limitations - **Sim-to-Real Gap:** Although verified for physical grounding, minor sim-to-real discrepancies may still occur in extremely high-precision contact mechanics tasks. - **Domain Scope:** While covering 8+ major scenarios, the dataset may not account for highly specialized or hazardous industrial environments not represented in the training distribution. ### Performance Benchmarks Empirical results demonstrate that the RoboGene pre-training policy significantly outperforms human-curated baselines across five unseen dual-arm tasks. Most notably, in the complex DFR-BuildBlocks scenario, RoboGene achieves a success rate five times higher than that of existing foundation models (25% vs. 5%). Furthermore, our approach exhibits superior generalization capabilities, maintaining robust performance in novel environments where baseline policies frequently fail to adapt. ## structure of the dataset Each task follows the LeRobot format: ``` task_name/ ├── data/ │ └── chunk-000/ │ ├── episode_000000.parquet │ ├── episode_000001.parquet │ └── ... ├── meta/ │ ├── info.json # │ ├── tasks.jsonl # task description │ ├── episodes.jsonl # │ └── episodes_stats.jsonl └── videos/ └── chunk-000/ ├── observation.rgb_images.camera_front/ ├── observation.rgb_images.camera_left/ ├── observation.rgb_images.camera_right/ ├── observation.rgb_images.camera_top/ ├── observation.rgb_images.camera_wrist_left/ └── observation.rgb_images.camera_wrist_right/ ``` ### camera configuration（six RGB cameras) - `camera_front` - `camera_left` - `camera_right` - `camera_top` - `camera_wrist_left` - `camera_wrist_right` ## Citation If you use this dataset in your research, please cite: **BibTeX:** ```latex @article{zhang2026robogene, title={RoboGene: Boosting VLA Pre-training via Diversity-Driven Agentic Framework for Real-World Task Generation}, author={Zhang, Yixue and Wu, Kun and Gao, Zhi and Zhao, Zhen and Ren, Pei and Xu, Zhiyuan and Liao, Fei and Wang, Xinhua and Fan, Shichao and Wu, Di and others}, journal={arXiv preprint arXiv:2602.16444}, year={2026} }