GenManip is a comprehensive robotics simulation suite built on NVIDIA Isaac Sim, designed for research in general robotic manipulation. It provides an integrated platform for data generation, benchmarking, and baseline development, offering a unified workflow from precision scene design to large-scale dataset creation.

This repository contains installation instructions, tutorials, documentation, example benchmarks, and references for all baseline methods.

GenManip supports the full workflow—from handcrafted scenes to procedurally generated large-scale datasets. Its streamlined toolchain allows you to easily build, customize, and share your own manipulation tasks.

The core concept is the GenManip Package: Install official or community benchmarks just like adding expansion packs to a game. Everything communicates through a black-box unified API so you can focus on model development without worrying about internal implementations.

GenManip strictly follows LeRobot GR00t data conventions, ensuring compatibility with modern training pipelines.

• 🔌 GenManip Package System

  Install or publish benchmark assets with a single command — expandable like game DLCs.

• 📊 Unified Benchmark Interface

  Includes GenManip Scaling Pick-and-Place, GenManip IROS Benchmark, and more. All benchmarks share one unified communication API, making model evaluation plug-and-play.

• 🧩 User-Friendly Docs & Config Templates

  Rich tutorials and configuration examples help you get started in minutes. You can create your own benchmark or data pipeline with just a few config edits.

• 🎨 Full-Stack Domain Randomization

  Randomize objects, layouts, lights, cameras, textures, rooms, enabling robust large-scale data generation.

• 🤖 Cross-Embodiment Support

  Works out of the box with:

  • Franka Panda + Panda Hand
  • Franka + Robotiq 2F-85
  • Aloha Split
  • Lift2

• 📐 Rule/Execution Set System

  Provides a structured syntax for defining task completion logic (top / left / right / front / back / in relations + logical composition). Compute the rules and generate data by execution set, result in photorealistic manipulation data.

• 🚀 Massive Parallel Execution

  Run thousands of Isaac Sim instances across multiple servers. Stress-tested to 1500 concurrent instances on 500× RTX 4090 (48GB) GPUs.

• 🏭 High-Performance Data Generation Pipeline

  Built on cuRobo + generalized oracle rules. Scales from single GPU to hundreds of GPUs.

• 🧱 Meta Object System

  Flexible scene composition and object substitution for scalable dataset/benchmark creation.

You can launch your first benchmark or data generation pipeline in minutes. Check out our tutorials for a step-by-step learning path — from basics to advanced usage.

👉 Full tutorials available at genmanip.com

For questions or collaborations, feel free to open an Issue or contact: 📧 [email protected]

See our user case at: overview/#example-use-cases

If you find our work useful, please cite:

@inproceedings{gao2025genmanip,
  title={GenManip: LLM-driven Simulation for Generalizable Instruction-Following Manipulation},
  author={Gao, Ning and Chen, Yilun and Yang, Shuai and Chen, Xinyi and Tian, Yang and Li, Hao and Huang, Haifeng and Wang, Hanqing and Wang, Tai and Pang, Jiangmiao},
  booktitle={CVPR},
  year={2025}
}

Know more about our CVPR paper version at branch archived/cvpr2025