Paper Review | RoboGen
RoboGen: Advancing Robot Design with AI-Powered Generative Models
Serendipity. The other day, while sitting on the train to Munich, I met this guy from RobCo, a Dresden-based Sequoia investment, who is working on modular industrial robots. While I don’t want to disclose here how long it takes for RobCo to train their robots, we started to talk about the complexities of training industrial robots in a variety of skills. Industrial robots are usually bought to do one thing and do this one thing well. Yet, we both agreed that there is always room for improvement in terms of speed, cost, and quality.
An innovative project that tries to tackle the problem of Skill Learning of Robotis is RoboGen. RoboGen is a joint project from researchers at CMU, Tsinghua IIIS, MIT CSAIL, UMass Amherst, and MIT-IBM AI Lab.
Goal: Train robots fast and inexpensive with a diverse set of skills, enabling them to operate in various settings, therefore helping them perform a broad range of tasks.
Problem: Training robots is expensive, time-consuming, and requires specialized materials and skills.
Solution: Learning skills through “Generative Simulation”. The robotic agent generates tasks and environments through foundation models and uses those to acquire skills autonomously.
Opinion: This paper was a sloth to get through, but it was worth it. While maybe for current industrial robots that are used in production chains, this project does not solve a real-world problem, I believe RoboGen still adds immense value. RoboGen aims to tackle the problems of employing Internet-scale data for Deep Learning by creating synthetic environments. Through this, it might help broaden the adoption of robots in general tasks.
So let’s dive in from the top.
