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

我们提出了RoboGen，这是一个生成式机器人代理，通过生成式模拟自动学习多样化的机器人技能。RoboGen利用了基础和生成模型的最新进展。我们提倡一种生成式方案，而不是直接使用或改编这些模型来产生策略或低级行动，该方案使用这些模型自动生成多样化的任务、场景和训练监督，从而在最小人类监督下扩展机器人技能学习。我们的方法为机器人代理配备了自我引导的提议-生成-学习循环：代理首先提出有趣的任务和技能来开发，然后通过填充相关物体和资产以及正确的空间配置来生成相应的模拟环境。之后，代理将提出的高级任务分解成子任务，选择最佳的学习方法（强化学习、运动规划或轨迹优化），生成所需的训练监督，然后学习获取所提出的技能的策略。我们的工作旨在提取嵌入在大规模模型中的广泛而多样化的知识，并将其转移至机器人领域。我们的完全生成式流水线可以重复查询，产生与不同任务和环境相关的无尽技能演示。

The paper aims to address the problem of scaling up robotic skill learning with minimal human supervision by leveraging generative simulation.

The key idea of the paper is to use generative models to automatically generate diversified tasks, scenes, and training supervisions, thereby enabling a self-guided propose-generate-learn cycle for the robotic agent.

The paper proposes a fully generative pipeline that can be queried repeatedly, producing an endless stream of skill demonstrations associated with diverse tasks and environments. The experiments were conducted on various robotic tasks, and the results showed that the proposed method outperformed the state-of-the-art methods in terms of sample efficiency and generalization. The paper also provides an open-source platform called RoboGen for generating robotic agents with diverse skills and environments.

Some related works in this field include 'Learning Dexterous In-Hand Manipulation' by OpenAI, 'Learning to Simulate Complex Physics with Graph Networks' by Google, and 'Sim-to-Real Transfer of Robotic Control with Dynamics Randomization' by UC Berkeley.