Pandemic Presents Opportunities for Robots; Teaching Them is a Challenge 

Keurig Dr Pepper is pursuing AI to fuel an expansion strategy as it commits to Google Cloud and enjoys a market share increase from coffee-drinkers. (Credit: Getty Images) 

By John P. Desmond, AI Trends Editor  

The pandemic is opening up opportunities for robots in the restaurant business as kitchens look for ways to distance employees and customers.  

White Castle, the regional hamburger restaurant chain, is testing a robot arm from Miso Robotics, for cooking french fries and other food, according to an account in the Associated Press. The two companies had been in discussions for about a year; talks picked up when the coronavirus hit. One potential benefit is the robot can free up time for human staff to handle increasing delivery orders.  

Vipin Jain, co-founder and CEO, Blended

Robot use by the restaurant industry is expected to pick up. “I expect in the next two years you will see pretty significant robotic adoption in the food space because of Covid,” stated Vipin Jain, the co-founder and CEO of Blendid, a Silicon Valley startup. 

Blendid’s robot kiosk makes fresh smoothies according to a recipe customers tweak from their smartphone app. Blendid employee keeps ingredients refilled once or twice a day. The company has a handful of kiosks operating around San Francisco, and is making sales outreaches to hospitals, shopping malls and supermarkets. “What used to be forward-thinking—last year, pre-Covid—has become current thinking,” Jain said. 

Max Elder is skeptical about the future of food prep robots. In the same AP article Elder, the research director of the Food Futures Lab at the Institute of the Future in Palo Alto, argued that food is personal and needs a human touch. He worries that robots will distract from important  human rights issues in the food industry including meat industry workers or produce pickers. “We can’t automate our way out of the pandemic because the pandemic affects much more than what can be automated,” he said. 

Beyond the question of how far robots can go to replace humans in the restaurant and food service industries, is the question of how robots can be trained to do the tasks required of restaurant workers or even to execute household chores. 

Robots Learning to Set the Table by Observing Humans 

Ankit Shah, graduate student, Department of Aeronautics and Astronautics, MIT

Robot researchers at MIT, for example, are studying ways robots can learn new tasks simply by observing humans, for instance setting the dinner table, according to a paper published in IEEE in February (DOI: 10.1109/LRA.2020.2977217) 

Researchers compiled a dataset with information about how eight objects—a mug, glass, spoon, fork, knife, dinner plate, small plate, and bowl—could be placed on a table in various configurations. The robot arm observed human demonstrations of setting the table with the objects, then the researchers asked the arm to set the table itself based on what it had seen. The robot made only a handful of mistakes in thousands of tests.  

“The vision is to put programming in the hands of domain experts, who can program robots through intuitive ways, rather than describing orders to an engineer to add to their code,” first author Ankit Shah, a graduate student in the Department of Aeronautics and Astronautics (AeroAstro) and the Interactive Robotics Groupsays in MIT Press coverage of the paper. Shah says the current paper reflects an early step in fulfilling the vision—giving robots the ability to learn without programming from ordinary workers. “Factory workers can teach a robot to do multiple complex assembly tasks. Domestic robots can learn how to stack cabinets, load the dishwasher, or set the table from people at home, he envisions.  

The researchers call the robot’s software system Planning with Uncertain Specification (PUnS), which enables a robot to hold a “belief” about the most likely correct outcome over a range of possibilities. “The robot is essentially hedging its bets on what is intended in a task, and takes actions that satisfy its belief, instead of us giving it a clear specification,” Shah said.  

The system is built on “linear temporal logic” (LTL), which can encode logic about future paths according to various time-based decisions.  The robot’s observations of 30 human demonstrations for setting the table yielded a probability distribution over 25 different LTL formulas. Each formula encoded a slightly different preference—or specification—for setting the table. That probability distribution becomes its belief.  

“Each formula encodes something different, but when the robot considers various combinations of all the templates, and tries to satisfy everything together, it ends up doing the right thing eventually,” Shah stated. 

The programming enables the robot to be flexible. If no forks are provided for the table-setting, the robot puts everything else in the right place. When the forks are introduced, the robot puts the forks in the right place. “That’s where flexibility is very important,” he stated. “Otherwise the robot would get stuck when it expects to place a fork and not finish the rest of the table setup.”  

Robots to Help Older Folks at Home Need Prior Knowledge and a Way to Learn 

Leslie Pack Kaelbling, Professor of Computer Science and Engineering, MIT CSAIL

Another MIT researcher is studying how to train robots to help older folks in their home.  

The robot would have to be shipped with a considerable amount of prior knowledge and ability, but it would also need to be able to learn on the job,” Leslie Pack Kaelbling of the Computer Science and AI Lab of MIT said in a recent article in Science. 

Richard Sutton, a pioneer of reinforcement learning, has asserted that humans should not try to build any prior knowledge into a learning system because we tend to get it wrong, Kaelbling noted. She and her team have turned to a “metalearning” approach, essentially learning about learning. They use machine learning early in system design to discover what structured algorithms and prior knowledge would enable the robot to learn efficiently when it is deployed in the wild.”  

A meta-learner tries to learn a learning algorithm that, when faced with a new task or environment in the wild, will learn as efficiently and effectively as possible.  

After careful consideration, the researchers decided that meta-learning will not fit the bill either. They intend to explore other options including teaching by humans, collaborative learning with other robots, and changing the robot hardware along with the software. “In all these cases, it remains important to design an effective methodology for developing robot software,” the researchers write. “Applying insights gained from computer science and engineering together with inspiration from cognitive neuroscience, can help to find algorithms and structures that can be built into learning agents and provide leverage to learning both in the factory and in the wild.” 

Teaching robots to perform useful tasks in restaurants and in the home is a work in progress. 

Read the source articles from the Associated Press, in MIT News and in  Science.