Brandon Araki started his academic journey as an undergraduate at Yale University studying biology, but his participation in an engineering research internship via the Harvard REU program at the end of his junior year helped guide his eventual future at CSAIL. Araki credits the internship for showing him how fulfilling engineering research can be and for convincing him to switch his major from biology to mechanical engineering. The internship coupled with the change in major put Araki in a position to do relevant research his senior year of college.
Upon graduation, Araki said he wanted to continue research and pursue a PhD. Araki said he was interested in robotics, so MIT was the best route considering the strength of the school’s robotics program. Araki ended up at CSAIL slightly by chance. After initially getting accepted to the mechanical engineering master’s program, he messaged several robotics professors in both mechanical engineering and CSAIL. Araki soon received an invitation to join Professor Daniela Rus’ lab.
Araki said he is excited about his current project “Learning and Planning with Logic.” Deep learning is facing a few big hurdles to real world deployment, including its need for large amounts of data and the difficulty of interpreting and explaining what it has learned, according to Araki. He said his project aims to address these problems, at least in the narrow scope of robotic task learning.
“We are using ideas and tools from formal logic, Bayesian inference, and reinforcement learning to design a system that can infer the rules governing a task (such as driving or cooking a dish) and learning how to execute the task itself from just a few example demonstrations,” Araki said. “Our work builds on a powerful idea – that it is possible to combine deep learning with probabilistic models to make algorithms that need less data, are more interpretable, and give deeper insight and better performance than what deep models or probabilistic models alone could do.”
While the work is probably a long way from real world use, Araki said that he hopes it contributes to the growing body of evidence that deep probabilistic models are an interpretable, data-efficient alternative to the current paradigm of deep learning.
Araki said working with Rus’ group has been a great experience given her range of interests. Their group has sub-groups working on autonomous vehicles, soft robotics, coresets and more. Araki said the group has exposed him to the full stack of robotics research from manufacturing to programming to algorithm development to theory.
“Since joining her group, I’ve worked on haptic belts to help blind people navigate; robot design for miniature flying-and-driving vehicles; multi-robot localization and path planning; algorithms for autonomous vehicles; and, finally, my current project, learning and planning with logic,” Araki said. “The projects have involved coming up to speed in many areas – embedded electronics, mechanical design and fabrication, robotics software, control theory, Bayesian inference, deep learning, and reinforcement learning.”
Throughout his time at MIT, Araki has also switched programs from the master’s in mechanical engineering to the PhD for EECS.
“As someone who took just one computer science class in college, it’s a transformation that I never could have anticipated,” Araki said. “I’m very grateful to Daniela for her faith in me, and I don’t think many groups could accommodate such a strange research path.”
After CSAIL, Araki is looking to find work either as a research scientist at a corporate lab or as an entrepreneur.
“At heart I am an engineer, and I would really love for something that I design to improve the lives of a lot of people,” Araki said. “MIT has given me the tools to work on a wide range of technical projects, and after CSAIL the challenge will be to find a way to apply that knowledge in a fulfilling way.”