Teaching the Next Generation of Robots
Dr. Sergey Levine discusses the latest advances in robotic learning.
Published February 13, 2018
While the 20th century was defined by machines programmed by people to perform specific, repetitive tasks, Sergey Levine, PhD, Assistant Professor, UC Berkley, Electrical Engineering and Computer Sciences wants the 21st century to be defined by robots capable of learning from their own past experience and performing multifaceted tasks. A researcher in the field of machine learning, he hopes to use algorithms and other learning techniques to teach robots to acquire greater autonomy that allows them to develop complex behavioral skills.
Research from Dr. Levine’s lab and other institutions has shown that robots can learn to successfully perform tasks like grasping objects through a system of trial and error or visualization of the task being performed. “A major goal of reinforcement learning and robotic learning research is to enable robots to autonomously learn how to perform a task. Someone still needs to program (or specify, or show) what the task actually is,” he explained. “In a sense, robots can already learn like humans because they can improve with experience. My research is concerned with making this a practical tool approach, improving how fast robots can learn, and proficiency at various tasks through autonomous learning.”
Dr. Levine’s approach to robotic learning differs from other successful machine learning strategies. “Many of the successes of machine learning in recent years use what is called supervised learning: a setting where the machine is provided with example inputs (e.g., images of objects) and their true labels (e.g., the category of the object),” he explained. Robotic learning is different. “The outputs are typically abstract and hard for people to specify manually, like joint angles or motor voltages, and the robots have to explore various options themselves to find the correct one. The level of supervision is much weaker, and active exploration of the environment is typically needed.”
This type of learning is not without its challenges. It is very difficult for a robot to visualize the world around them because “the physical world is highly varied and often unpredictable,” said Dr. Levine. “The difficulty really comes from the diversity and breadth of the real world and the range of different tasks that a truly ‘generalist’ robot would need to accomplish.” He acknowledged that, “what makes human learning so incredibly powerful is not that humans are particularly excellent at any one thing, but that they are so adaptable as to be able to do pretty much anything, if given enough practice.”
Yet Dr. Levine is optimistic that these challenges will be overcome and that researchers are on the cusp of making breakthroughs in the service or industrial sector that will positively impact people. He anticipates that in the next five years, robotics potential will be able to automate a wide range of physical tasks that right now are routinely the province of humans, such as eldercare or care for people with disabilities.
This optimism is belied by a popular culture that sees the rise of robots as an alarming development. In particular Dr. Levine is encouraged by more recent positive portrayals of machines and robots in media. “Big Hero 6 is an excellent example that I like very much – an illustration of how technology, artificial intelligence, and the scientists who work on it can help make the world a better place,” he said. “I do however think that there is also cause for caution when it comes to robotics and artificial intelligence, in the same way as we should be cautious about any powerful new technology. We should be cognizant of the dangers and make sure that we as a society use technology responsibly and ethically.”
Hear more from Sergey Levine, PhD via Livestream when he appears at the Academy on Friday, March 9th 2018 2018 at 4:05pm for the 12th Annual Machine Learning Symposium to present a keynote lecture titled “Machines that Learn by Doing.” A second keynote lecture titled “Data Genesis: Examining and Accounting for the Data that Trains AI Systems” presented by Meredith Whittaker of AI Now Institute and Google is also scheduled.