Deep learning-driven robotic systems are bottlenecked by data collection: it’s extremely costly to obtain the hundreds of thousands of images needed to train the perception system alone. It’s cheap to generate simulated data, but simulations diverge enough from reality that people typically retrain models from scratch when moving to the physical world.
We’ve shown that domain randomization, an existing idea for making detectors trained on simulated images transfer to real images, works well for cluttered scenes. The method is simple: we randomly vary colors, textures, lighting conditions, and camera settings in simulated scenes. The resulting dataset is sufficiently variable to allow a deep neural network trained on it to generalize to reality.
The detector is a neural network based on the VGG16 architecture that predicts the precise 3-D location of Spam in simulated images. Though it has only been trained on simulated scenes, the resulting network is able to detect Spam in real images, even in the presence of never-before-seen “distractor” items arranged in random configurations.
The video below demonstrates the system in action: