Post by Toyota Research Institute

How can we solve the robotics data gap? Ken Goldberg, Professor at UC Berkeley, sits down with Richard Cheng, Senior Robotics Research Scientist at TRI, to discuss the Real2Render2Real (R2R2R) project, a collaborative effort to develop a scalable pipeline that synthesizes thousands of training examples from just one human demonstration. Some takeaways: • The amount of data we have for robot training is 100,000x less than what exists to train vision-language models • R2R2R can generate diverse 3D training scenarios without intensive human or robot labor • Policies trained this way perform as well as those using 150 manual human trials Learn more about this University Research Program project: https://lnkd.in/gQTJ9KY9

Video Content