Turning daydreams of simulation into scalable synthetic data — Announcing Compound’s investment in AI.Reverie
A few years ago while spending time with autonomous vehicle (AV) engineering teams I kept asking what their largest bottlenecks were to improving autonomy. One constant I kept hearing was more annotated training data. The bottleneck of taking sensor feeds and having humans annotate them in order to train deep learning perception models was resulting in teams of hundreds and even thousands of humans performing this time-consuming task.
At the same time, I started to learn more about simulation and the value that synthetic data could bring to this pipeline. I wrote about this extensively here and set out to make an investment in the space with the thesis that synthetic data would be a key accelerator in the arms race towards autonomy for AVs, and that the addressable market for other robots and autonomous systems would expand over time.
I’ve said this before, but being thesis driven allows us to move efficiently within categories we’ve gone deep on, but also makes it even more powerful when we can recognize when founders shatter our theses.
After meeting with 10+ companies from pure simulation to larger scale world building, all related to autonomous vehicles, I finally met Daeil Kim & Paul Walborsky. They introduced me to AI.Reverie. My daydreams of all the possibilities of synthetic data had turned into a reality.
While AV simulation technology continues to be extremely valuable, Daeil and Paul, along with the rest of the AI.Reverie team have built a simulation platform that creates large scale, high-fidelity worlds algorithmically in order to train machine learning models across markets from more traditional enterprise robotics all the way to tracking elephants. Not only was I blown away by the scalability of what they had built, but also the level of diversity at which they could generate virtual worlds based off of real location data.
Their vision for the future recognized the value of fully annotated, error-free synthetic data and simulation across use-cases I had not seen any other company even conceptualize, let alone commercialize.
We’re super excited to be backing the entire AI.Reverie team in their pursuit of helping train AI to understand the world.