Research map / Vision-language and multimodal / Robot action and navigation
IGen: Scalable Data Generation for Robot Learning from Open-World Images
Chenghao Gu, Haolan Kang, Junchao Lin, Jinghe Wang, Duo Wu, Shuzhao Xie, Fanding Huang, Junchen Ge, Ziyang Gong, Letian Li, Hongying Zheng, Changwei Lv and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Real-to-Sim digital-twin reconstruction with IGen scene reconstruction from a single image | scene reconstruction · Simpler scenes, average of 4 tasks | PSNR: 17.2649 -> 27.0040 | Improved |
| replace Real-to-Sim with IGen scene reconstruction | scene reconstruction · Simpler scenes, average of 4 tasks | SSIM: 0.6833 -> 0.8522 | Improved |
| replace Real-to-Sim with IGen scene reconstruction | scene reconstruction · Simpler scenes, average of 4 tasks | LPIPS (AlexNet): 0.3235 -> 0.0630 | Improved |
| use IGen instead of Cosmos-Predict2 and TesserAct for behavior generation | robot behavior video generation · DreamGen Bench, Instruction Following and Physics Alignment | proportion of videos scored 1: nearly twice as many Instruction Following successes with Qwen-3-VL-Plus | Improved |
| use IGen instead of TesserAct and Cosmos-Predict2 for data generation | data generation efficiency · per-sample generation under same GPU memory | efficiency: approximately 30x and 200x higher efficiency | Improved |
| fine-tune pi0 on IGen-generated data instead of no fine-tuning | pi0 · real-world Place Bottle task | success rate (%): 0.0% -> 75.0% | Improved |
| use 100 IGen-generated samples instead of 100 teleoperated samples (same time budget) | pi0 fine-tuning · real-world manipulation tasks | success rate (%): 19.4% -> 44.5% | Improved |
| scale IGen-generated data from 100 to 1,000 samples (ablation) | pi0 fine-tuning · real-world manipulation tasks | success rate (%): 44.5% -> 66.7% | Improved |
| fine-tune pi0 on 1,000 IGen samples instead of teleoperation data under same collection time | pi0 fine-tuning · real-world manipulation tasks | success rate (%): 58.3% -> 66.7% | Improved |
Closest papers by meaning
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.90)
- Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation (ICLR 2024 · similarity 0.89)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.89)
- What Matters in Learning from Large-Scale Datasets for Robot Manipulation (ICLR 2025 · similarity 0.88)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.88)
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
