Robots, machine learning, and self-driving cars were big topics at the recent White House Frontiers Conference. While we’re living in a time of great innovation and change, what could it mean for humanity if robots could learn from trial and error?
To explore this question, Pieter Abbeel, an Associate Professor at UC Berkeley, spoke to a group of early career researchers and science writers about his latest research in machine learning, the significance of the results, and potential next steps for the research. The early career researchers in attendance were mostly UC Berkeley students interested in machine learning and robotics. The event, which took place at the Z Café in Oakland this month, was one of the quarterly networking dinners hosted by the Northern California Science Writers Association (NCSWA). The NCSWA dinners offer a great opportunity for ECRs to network and learn about new research in interesting and emerging fields of study.
Abbeel began his presentation with a video of a robot slowly folding a shirt fresh from the laundry. Robots can be programmed to execute rote tasks, such as folding a shirt, by following a specific set of instructions. However, Abbeel and his colleagues are interested in whether a robot can learn to perform tasks, and expand its repertoire beyond performing the same action over and over. The robot may be able to reliably fold a shirt of the same size if it is placed on the table in precisely the same way, but what happens if it encounters a pair of pants?
Now that there is a greater amount of data available and stronger computing power than ever before, it is possible to enable a robot to learn through trial and error. Abbeel explained how a set of algorithms could be used to model higher-level abstractions in robots, otherwise known as deep learning.
While machine learning can help make our lives easier, there are many important considerations and improvements to the technology to be made. Some potential next steps for this research are enabling a robot to share a skill with another robot, and to enable a robot to think abstractly to set goals for the future, both which are abilities currently restricted to humans.
If you are interested in learning more about Pieter Abbeel’s research, or the subject of machine learning in general, there are a couple of book recommendations from him below. Additionally, if you’re interested in attending future NCSWA events and networking with fellow ECRs and potential science graduate students, sign up for a student membership here.
The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies by
Erik Brynjolfsson and Andrew McAfee
Superintelligence: Paths, Dangers, Strategies by Nick Bostrom
Image Credit: Tessa Gregory