# RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation

**Authors:** Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, Chuang Gan  
**arXiv:** [2311.01455](https://arxiv.org/abs/2311.01455)  
**Date:** 2023-11-02  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2311.01455v1/

## Key points

- RoboGen is a generative robotic agent designed to acquire diverse skills autonomously through generative simulation.
- It leverages foundational and generative models to automatically create varied tasks, scenes, and training supervisions.
- The core concept involves a self-guided propose-generate-learn cycle where the agent proposes tasks, generates simulation environments, breaks down tasks into sub-tasks, selects optimal learning approaches, generates training supervision, and learns policies.
- The goal is to tap into large-scale models' knowledge and transfer it effectively to robotics while minimizing human involvement.
- The fully generative pipeline can produce an endless stream of skill demonstrations associated with diverse tasks and environments.

## Summary

In their paper titled "RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation," authors Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan introduce RoboGen - a cutting-edge generative robotic agent designed to autonomously acquire diverse robotic skills on a large scale through generative simulation. By harnessing the latest advancements in foundational and generative models, RoboGen takes a unique approach by not directly utilizing or adapting these models to generate policies or low-level actions. Instead, the authors advocate for a generative scheme that leverages these models to automatically create varied tasks, scenes and training supervisions. This innovative method enables the scaling up of robotic skill learning with minimal human intervention. The core concept behind RoboGen is to equip a robotic agent with a self-guided propose-generate-learn cycle. Initially, the agent proposes intriguing tasks and skills to develop before generating corresponding simulation environments by populating relevant objects and assets with appropriate spatial configurations. Subsequently,the agent breaks down the proposed high-level task into sub-tasks and selects the optimal learning approach (such as reinforcement learning,motion planning or trajectory optimization). It then generates necessary training supervision and learns policies to master the proposed skill. The overarching goal of this work is to tap into the vast and versatile knowledge embedded within large-scale models and transfer it effectively to the realm of robotics. The fully generative pipeline developed by RoboGen can be queried repeatedly to produce an endless stream of skill demonstrations associated with diverse tasks and environments. Through this groundbreaking research effort,the authors pave the way for automated robot learning at an unprecedented scale while minimizing human involvement in the process.