GenManip v1.0.0 Pre-release

We’re excited to announce GenManip Suite v1.0.0-prerelease, a comprehensive simulation platform for data generation and benchmarking, built upon NVIDIA Isaac Sim. GenManip Suite provides high-quality physics simulation and rendering, with a powerful GUI that allows users to easily configure data generation tasks and benchmarking pipelines. It currently supports multiple robot embodiments — including Franka PandaHand, Franka Robotiq, Aloha, and Lift2 — and offers convenient configuration through YAML files with extensive domain randomization.

Visit our bilingual website for detailed guidance: https://genmanip.axi404.top/

GenManip originated from our CVPR 2024 paper,
GenManip: LLM-driven Simulation for Generalizable Instruction-Following Manipulation. Since then, our focus has shifted toward data generation and evaluation, building upon the original codebase to form today’s GenManip Suite. To date, GenManip has demonstrated strong potential as a scalable data engine. For example, the InternData-M1 dataset was generated with GenManip across tens of thousands of assets, and its multi-embodiment and long-horizon variants are already complete and ready for release. In collaboration with InternManip, GenManip-powered datasets and benchmarks have enabled the IROS 2025 Challenge of Multimodal Robot Learning in InternUtopia and Real World, showcasing its potential for well-designed and scalable benchmarks. Many other collaborations are underway — exploring articulated object manipulation, long-horizon scaling, and integration with next-generation scene generators — all of which will gradually be integrated into the GenManip community.
We warmly welcome potential collaborators to get in touch!

🚀 What’s New in GenManip v1.0.0-prerelease

• Extensible data generation — easily extend tasks to longer horizons through simple goal duplication and task definition via scene graphs
• Out-of-the-box evaluation — distributed design and clear interface standards allow users to run benchmarks without touching low-level simulation code; you can also define your own tasks
• Fine-grained layout control — design layouts directly in the Isaac GUI or specify object placement and randomization via configuration files
• User-friendly codebase and documentation — set up your own generation or evaluation task in as little as 7 minutes
• Scaling-up support — efficient object pool management, multi-server parallelization, and robust handling of large-scale data generation
• Support for multiple embodiments — including Franka, Aloha, and Lift2
• Extensive domain randomization — lighting, texture, and initialization diversity

🔮 What’s Coming Next

We’re actively exploring the boundaries of simulation-based data generation:

• Articulated asset support — rule-based data generation for articulated objects without manual labeling
• More robot embodiments — expanding support for a broader range of robot types

Our documentation is continuously being updated — all code is available here, and the complete documentation will be finalized within a week.

We welcome issues, suggestions, and collaboration inquiries. For research collaborations or technical support, please contact: [email protected]