Startup Blendid is trekking into the future with smoothie-making robots.
The Silicon Valley robotics company is using AI powered by the compact supercomputing of NVIDIA Jetson to quickly serve up customizable juice and vegetable blends.
Co-founder Vipin Jain said the original Star Trek was what inspired him to develop Blendid.
“Star Trek’s replicator was an intelligent machine that knew you, what you liked to eat and how you liked it prepared. That is what I want with Blendid,” said Jain, the company’s CEO.
But that’s just half the story.
Co-founder Venki Ayalur, the company’s CTO, has long been interested in digitizing recipes in order to preserve memories of preparing food. He strives to create healthier and tastier food options than what’s often available today.
Together with their team, they’ve developed robots that can autonomously make smoothies with creative ingredients.
Blendid’s smoothie robots are now installed at the University of San Francisco campus, at a cafeteria in a high-tech company in the Bay Area, and at Plug and Play Tech Center in Sunnyvale. The founders have plans for a few more locations this year, including a spot in downtown Palo Alto near Stanford University.
Blendid’s executive chef, Kristen Rasmussen de Vasquez, a nutritionist, created the recipes the bots whip up.
She teaches food science at the University of California, Berkeley, and focuses on the benefits of natural ingredients.
Her smoothie recipes include items like flax and chia seeds, kale, cinnamon, cocoa, ginger and the superfood matcha for their health benefits. Customers can customize the ingredient choices before placing their order.
After a customer selects optional ingredients, the bot automatically adjusts the other ingredients so that the smoothie has good taste and texture and remains the right serving size.
By requiring its robots to rely on ingredient servings by weight, Blendid can deliver smoothies that are made precisely to recipes and customizations.
Blendid’s robot smoothie baristas make it look easy, but there’s actually a lot going on. The system is powered by the compact supercomputing of Jetson TX2, allowing it to ingest data from its nearly two dozen sensors, including cameras for image recognition.
Data scientists at Blendid gathered hundreds of images for each ingredient to train deep neural networks to accurately spot them. They also used transfer learning and the ImageNet database, as well as MobileNet COCO and Faster R-CNN neural networks to refine the results.
The system is intelligent. Its image classification networks are able to adjust the robot’s motions to reach for ingredients misplaced by workers servicing the robotic juice stations.
And if you’re looking for fast, it can complete 45 smoothies per hour while handling as many as nine orders at a time.
NVIDIA GPUs on Google Cloud Platform enabled Blendid to accelerate the training of its neural networks.
“If you have good data and insights that represent your field environment, you can get to 99 percent or better accuracy with your vision solution,” Jain said.