//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
A group of researchers developed a machine learning (ML) technique for controlling a personal robot that leads to better performance with lower data input. The technique helps a non-technical robot owner figure out why a bot failed in a task—and then correct it themselves instead of shipping it back to the factory.
“In this work, we take the perspective that different users may want very different things done in their home,” Andi Peng, an MIT graduate student in electrical engineering and computer science (EECS), told EE Times. “But what they fundamentally want in terms of the tasks is not that different from something the robot may already know. So, the question is, how do you extract that extra knowledge required to improve what the robot already knows until it’s doing what the human wants?”
Peng and a team from MIT, the Stevens Institute of Technology, and the University of California, Berkeley, developed an algorithm and used it in simulation to ask the robot’s human owner for information after the bot failed a task, figure out what the gap was in the robot’s knowledge, and take steps to fix it. Their research used a simulation app, a visual motor attention agent (VIMA), that Stanford University developed.
The researchers used a Universal Robots UR5e stationary, collaborative robot (or cobot) in their simulation.
“The problem is, if every time you need a new task done and you have to redo this process [training the robot], then you’re wiping the memory off the robot and then teaching it something entirely new—so there’s no continual adaptation,” she said. “Instead of training from scratch with new data, we find a way to adapt the current algorithm using machine learning in a faster way.”
For example, if the cobot were trained to pick up a red book, it may fail if directed to pick up a book that’s blue. In that instance, the team’s system uses an algorithm to create “counterfactual” explanations that identify what needs to change for the robot to succeed. It then gets feedback from the human about why the robot failed, and uses the feedback and counterfactual explanations to generate new data to fine-tune the bot.
“We literally output a demonstration of the robot doing the task correctly in a counterfactual situation,” Peng said. “And, from that demonstration, we can basically perform what’s called an augmentation process. Then, with that, we can kind of buy new data for free.”
The robot can then pick up a book of any color without having been trained on thousands of volumes. The alternative would be to send the bot back to the factory for retraining from scratch.
In July, Peng and fellow MIT EECS grad student Aviv Netanyahu, a co-collaborator on the research, showed the results of their work in a poster presentation at the 40th International Conference on Machine Learning.
Robot training 1.1
Netanyahu explained the system works only for objects that can be picked up in a similar way. Their simulation was with a suction attachment on the end of the cobot’s arm trained to pick up common items found in a home like frying pans, boxes and toy blocks. It would work with neither an object it wasn’t trained on nor a different end-of-arm attachment like a gripper, she said.
“If you need to grasp something in a very, very different way, then we can’t just change the color and apply the same actions and get the robot to work,” Netanyaho said. “We would need new actions. So, in that case, we could adapt based on what the human is giving us, but we wouldn’t be able to use all the training information we had because those were using different actions. So that’s maybe 1.1.”
Putting ‘person’ in ‘personal bot’
Peng and Netanyahu’s work is all about human-robot interactions.
“We’re motivated by the idea that the end user in a home or somewhere else is the person that we need to tailor specific algorithms to,” Peng said.
If personal robots are to become more prevalent, she and her fellow researchers need to expand the demographic of bot users beyond the tech savvy.
To do that, they develop bot-control methods for older, non-techies, as well as people with disabilities.
“We really wanted to explore what happens when you have this distribution shift, what happens when your home is suddenly very different than your factory” where a robot is trained, Netanyahu said. “And that is the main thing we’re pushing, or trying to research: What happens when you have the same tasks, but things change? You still want your robot to work.”