Creating a blended drink, such as a fruit or vegetable smoothie, is one of those kitchen tasks that you have to be really motivated to complete. Between the cutting, squeezing and blending that all go on through the process, it’s one of those areas ripe for robots and automation. Food robotics company Blendid is taking this challenge head-on, through a combination of software, hardware, and advanced robotics to create healthy and delicious smoothies it calls “Blends.”
The company’s stand-alone kiosk consists of refrigeration systems, blenders, robotic arms, and dispensers. Customers can place an order in person at a kiosk, or through the Blendid app, at a cost of $6 for a 12-ounce drink. The company opened its first kiosk at the University of San Francisco’s main dining hall, Market Café, earlier this year, as well as at the Plug and Play Tech Center in Sunnyvale, Calif. Further deployments are scheduled as well, the company said.
Vipin Jain, Blendid CEO
Robotics Business Review recently spoke with the company’s CEO, Vipin Jain, about the technology used to create the kiosks, and the trends that allowed the company to develop its system.
Hybrid approach to AI
Q: Talk about how the artificial intelligence and machine learning is being applied to the system. It looks like the application and combination of ingredients are just part of customization options for customers. Is that where the AI/ML lies, or is it in some other part of the system?
Jain: There are three areas where ML/AI should be applied when it comes to food automation. The first area is recipe adaptation, based on consumer taste/feedback and ingredient dispensing. Our system automatically under- or over-compensates all the ingredients within the bounds defined by a chef to ensure optimum taste and consistency.
The second area is machine vision for self-configuration, auto-calibration and error correction, which allows for a resilient and operationally flexible environment for food robotics. Both of these are already in play. The third area is in consumer recommendations, based on their taste profile.
Q: Is the computer vision and sensors used to direct the robot arm to grab the ingredients correctly, as part of the mixing process, or more for when pouring the final blend into a cup? Or is it a case where all of that motion/movement is programmed by the system? How precise or imprecise is the system in terms of leeway so the robot doesn’t pour the smoothie incorrectly?
AI and machine learning helps the robot arm know where to pour the smoothie correctly. Image: Blendid
Jain: It’s both – pre-programming and use of computer vision. For the sake of speed, reliability and cost, in a closed loop system, it is prudent to use a hybrid approach. You program the regions, and then use computer vision for fine tuning and ongoing calibration and changes.
We have achieved 99% accuracy right now. The rest is about feeding in corner cases to our training set based on our field deployments / experience and improving the outputs further.
Q: Since opening the first location in March, has the company done any other deployments?
Jain: We have two commercial deployments outside of Blendid headquarters. The first one is at the University of San Francisco in San Francisco. The second is at the Plug and Play Tech Center in Sunnyvale. There are more in the pipeline that will go live through rest of this year. Interest from both consumers and potential operators has been tremendous.
Q: What lessons have you learned so far from the initial deployment? Were there any unexpected surprises or challenges “in the real world?”
Jain: Plenty. One can do all the lab testing, but a field environment is a different beast. How consumers interact with your product through various times of a day and the demand curve, you can only understand with a field deployment.
How operators handle servicing and error handling is not replicable in a lab environment. We are taking all this learning and continuously improving the product, the experience and operations.
Decreasing costs of robotics components have allowed for the creation of companies such as Blendid.
Q: What parts of the system have made it achievable to create this system? Robot arm development, lower costs of manipulation; advances in AI/ML and software, etc., or grippers/sensors advancing? We know you started this in 2015, what was it about that year that made you go, “Ah, robot smoothies!”
Jain: Robot smoothies wasn’t an “aha” moment. The realization that robotics and AI had gotten to a point where it could be used to build a system for food automation (like the Star Trek replicator) was that “aha” moment. From there, we quickly zoomed into the “smoothies / blends” as the beachhead, a trendy format that was big and complex enough to automate. It was still a tough road from there to Blendid, with four internal design and prototypes before we built the current Blendid system.
Q: We’re seeing many companies building systems based on “off the shelf” components, such as robot arms by Universal Robots and other cobot makers. How did the lower cost here, and in other areas (for example, development of AI/ML algorithms based on chip development or cloud data availability) drive the development of the complete application?
Jain: The cost of components and processing always results in new innovation. It was cost prohibitive to build a system for food automation five years ago, but not anymore. The use of standardized components, availability of ML/AI libraries, and cost-effective processing and storage in the cloud has reduced the cost of development and support significantly.
You can look at numerous innovations benefiting from these trends. Besides food automation, autonomous driving, delivery robots, and drones are all good examples.
Q: Why does the company create healthy blends instead of creating something like customized milkshakes?
Jain: Blends have a universal appeal, and are trendy across age groups and at different times of day (breakfast, lunch and afternoon options). And they are complex enough in terms of variety of ingredients that need to be handled with precision (across solids, liquids and super-foods) that we were confident that if we were able to automate blends, we would be able to automate various other formats and cuisines over time.
We didn’t want to build a business around a one-trick pony, but build a platform for food automation.
Q: But this system could be adapted to do other beverages, such as milkshakes, coffee, tea, etc.?
Jain: Of course, but these are trivial compared to blends. Our ambition is much larger around food automation.
