Robotics work heralds ‘discovery factories’

You might say it began with teaching a robot to cook.

In a presentation to faculty members from many University departments, Carolina computer scientist Ron Alterovitz said that his research group taught a robot to perform cooking actions like scooping and transporting ingredients.

(L-R) Jim Cahoon, chair of the chemistry department, Ron Alterovitz, distinguished professor in the computer science department, and Angelos Angelopoulos, a fourth-year graduate student, in the lab with the robot. (Johnny Andrews/UNC-Chapel Hill)

Alex Tropsha, the K.H. Lee Distinguished Professor in UNC’s Eshelman School of Pharmacy, told Alterovitz that chemistry researchers perform many lab tasks that resemble a sophisticated form of cooking.

The comment inspired a “what if” moment for Alterovitz, the computer science department’s Lawrence Grossberg Distinguished Professor. He saw the potential to transform labs in which humans handle tedious, manual tasks into automated discovery factories that make breakthroughs more quickly, safely and precisely.

“What if robots could maneuver freely in scientific labs to sense where humans are and adapt to changes?” Alterovitz wondered. “What if they could safely and autonomously fetch chemicals, pour and stir liquids and powders, operate equipment and do other processes?”

To answer those questions, the two collaborated with James Cahoon, professor and chair of Carolina’s chemistry department. They secured a Creativity Hubs grant from the Office of the Vice Chancellor for Research to fund the work in 2020.

The team wrote AI software that directs the actions of a one-armed robot on wheels. They experimented with the robot trying a common lab task — needle injection — that requires precise transfer of liquids and gases between instruments or stations.

“In robotics, it’s a challenge for a mobile robot to move multiple meters, then perform a manipulative task that requires submillimeter-level accuracy,” Alterovitz said.

But they did it.

The robot now glides across Cahoon’s lab carrying a sample, using sensors and cameras to find a gas chromatography machine. It unfolds its arm, moves closer. The AI software computes exactly where the arm must insert the needle to deposit its sample in the machine.

The scientists want to link tasks to automate the cycle — design a material or molecular system, make or synthesize it, test it, then analyze it. A new cycle then begins with an improved version of the material or molecular system based on the previous cycle’s analysis. Alterovitz notes, “Automating that design-make-test-analyze loop could have huge benefits by making the lab self-driving as it improves the material or molecular system over time.

“Robots are the quintessential application of AI, where we apply AI in the real physical world. Unlike a lot of AI forms used to create a picture or have a dialogue, the robot’s AI modifies and changes the physical state of the world,” said Alterovitz.

The automaton robot can handle tedious, manual tasks like transferring a syringe while maneuvering throughout the lab. (Johnny Andrews/UNC-Chapel Hill)

First-ever definitions

The robot’s the star. However, a recent Science Robotics paper published by Alterovitz and Cahoon outlines five definitions of automation levels that will widely influence how robotics can accelerate discoveries. The levels range from assistive automation with robots handling a simple, repetitive task to full automation, where robots automate multistep sequences and manage everything from setup to handling issues like equipment malfunctions.

“Defining automation levels is saying, ‘Here’s where we are and here’s where we want to go,’” Alterovitz said.

Autonomy is fundamentally different between robots in a mixed-use lab and robots in, say, a facility that produces cars or food products. In those factories, the whole environment is designed for the robot. The car moves on a track. Conveyor belts bring parts to robots. Fences surround robots. That won’t work in a lab where people move things.

“We’re focusing first on correctness and safety,” Alterovitz said. “Once those are in place, we’ll speed it up and it can work 24/7 in ways that humans don’t necessarily want to.”