<p>Today, perhaps, robotics is experiencing its ImageNet moment (the release of a large open dataset with images).</p>
<p>RT-X is the largest open dataset for robots ever assembled, including 33 institutes, 22 robotic devices, 527 skills, and 1 million episodes.</p>
<p>Why is robotics lagging behind NLP, computer vision, and other areas of AI? The main culprit is the lack of data, among other challenges. Unlike texts, images, and videos, you can't download large volumes of data on robot control from the internet. They simply don't exist.</p>
<p>11 years ago, ImageNet started a revolution in deep learning. 3-4 years ago, internet-scale data became the fuel for the first GPT and Diffusion models. I believe that 2023 will finally be the year of scaling robotics.</p>
<p>Major robotic models, such as VIMA (my team's work at NVIDIA) and RT-1/2 (Google DeepMind's project), are very
· Essay · 1 min