Teaching robots to teach other robots

author avatar

08 Jul, 2019

Optimus’ learned skills could be seamlessly transferred to Atlas, CSAIL’s 6-foot-tall, 400-pound humanoid robot.

Optimus’ learned skills could be seamlessly transferred to Atlas, CSAIL’s 6-foot-tall, 400-pound humanoid robot.

Most robots are programmed using one of two methods: learning from demonstration, in which they watch a task being done and then replicate it, or via motion-planning techniques such as optimization or sampling, which require a programmer to explicitly specify a task’s goals and constraints.

Both methods have drawbacks. Robots that learn from demonstration  can’t easily transfer one skill they’ve learned to another situation and  remain accurate. On the other hand, motion planning systems that use  sampling or optimization can adapt to these changes but are  time-consuming, since they usually have to be hand-coded by expert  programmers.

Researchers from MIT’s Computer Science and Artificial Intelligence  Laboratory (CSAIL) have recently developed a system that aims to bridge  the two techniques: C-LEARN, which allows noncoders to teach robots a  range of tasks simply by providing some information about how objects  are typically manipulated and then showing the robot a single demo of  the task.

Importantly, this enables users to teach robots skills that can be  automatically transferred to other robots that have different ways of  moving — a key time- and cost-saving measure for companies that want a  range of robots to perform similar actions.

“By combining the intuitiveness of learning from demonstration with  the precision of motion-planning algorithms, this approach can help  robots do new types of tasks that they haven’t been able to learn  before, like multistep assembly using both of their arms,” says Claudia  Pérez-D’Arpino, a PhD student who wrote a paper on C-LEARN with MIT Professor Julie Shah.

The team tested the system on Optimus, a new two-armed robot designed  for bomb disposal that they programmed to perform tasks such as opening  doors, transporting objects, and extracting objects from containers. In  simulations they showed that Optimus’ learned skills could be  seamlessly transferred to Atlas, CSAIL’s 6-foot-tall, 400-pound humanoid robot.

A paper describing C-LEARN was recently accepted to the IEEE  International Conference on Robotics and Automation (ICRA), which takes  place May 29 to June 3 in Singapore.

How it works

With C-LEARN the user first gives the robot a knowledge base of  information on how to reach and grasp various objects that have  different constraints. (The C in C-LEARN stands for “constraints.”) For  example, a tire and a steering wheel have similar shapes, but to attach  them to a car, the robot has to configure its arms differently to move  them. The knowledge base contains the information needed for the robot  to do that.

The operator then uses a 3-D interface to show the robot a single  demonstration of the specific task, which is represented by a sequence  of relevant moments known as “keyframes.” By matching these keyframes to  the different situations in the knowledge base, the robot can  automatically suggest motion plans for the operator to approve or edit  as needed.

“This approach is actually very similar to how humans learn in terms  of seeing how something’s done and connecting it to what we already know  about the world,” says Pérez-D’Arpino. “We can’t magically learn from a  single demonstration, so we take new information and match it to  previous knowledge about our environment.”

One challenge was that existing constraints that could be learned  from demonstrations weren’t accurate enough to enable robots to  precisely manipulate objects. To overcome that, the researchers  developed constraints inspired by computer-aided design (CAD) programs  that can tell the robot if its hands should be parallel or perpendicular  to the objects it is interacting with.

The team also showed that the robot performed even better when it  collaborated with humans. While the robot successfully executed tasks  87.5 percent of the time on its own, it did so 100 percent of the time  when it had an operator that could correct minor errors related to the  robot’s occasional inaccurate sensor measurements.

“Having a knowledge base is fairly common, but what’s not common is  integrating it with learning from demonstration,” says Dmitry Berenson,  an assistant professor of computer science at the University of Michigan  who was not involved in the research. “That’s very helpful, because if  you are dealing with the same objects over and over again, you don't  want to then have to start from scratch to teach the robot every new  task.”


The system is part of a larger wave of research focused on making  learning-from-demonstration approaches more adaptive. If you’re a robot  that has learned to take an object out of a tube from a demonstration,  you might not be able to do it if there’s an obstacle in the way that  requires you to move your arm differently. However, a robot trained with  C-LEARN can do this, because it does not learn one specific way to  perform the action.

“It’s good for the field that we're moving away from directly  imitating motion, toward actually trying to infer the principles behind  the motion,” Berenson says. “By using these learned constraints in a  motion planner, we can make systems that are far more flexible than  those which just try to mimic what's being demonstrated"

Shah says that advanced LfD methods could prove important in  time-sensitive scenarios such as bomb disposal and disaster response,  where robots are currently tele-operated at the level of individual  joint movements.

“Something as simple as picking up a box could take 20-30 minutes,  which is significant for an emergency situation,” says Pérez-D’Arpino.

C-LEARN can’t yet handle certain advanced tasks, such as avoiding  collisions or planning for different step sequences for a given task.  But the team is hopeful that incorporating more insights from human  learning will give robots an even wider range of physical capabilities.

“Traditional programming of robots in real-world scenarios is  difficult, tedious, and requires a lot of domain knowledge,” says Shah.  “It would be much more effective if we could train them more like how we  train people: by giving them some basic knowledge and a single  demonstration. This is an exciting step toward teaching robots to  perform complex multiarm and multistep tasks necessary for assembly  manufacturing and ship or aircraft maintenance.”

More by Adam Conner Simons

I'm a seasoned communications professional and writer/editor with more than a decade of experience in journalism and public relations, with particular expertise in media outreach and social media strategy.

Wevolver 2022