Chelsea Finn on Meta Learning & Model Based Reinforcement Learning

Memorable Moments

Chelsea Finn shares some memorable incidents from her work with robots, highlighting the unpredictable nature of robotics research. She recalls a humorous episode with a PR2 robot that, instead of completing its task, gestured for her to do it, showcasing a "lazy" robot behavior 1. Finn also recounts a more serious incident at Google, where an older robot's shoulder broke, yet it continued to operate, defying expectations of a shutdown 1. These stories illustrate the challenges and unexpected moments in robotics research.

That was not a lazy robot.

--- Chelsea Finn

Finn's experiences underscore the importance of supervision and adaptability in managing robotic experiments 2.

   

Experimental Setbacks

Experimental setbacks are a common theme in Chelsea Finn's robotics research, often due to logistical and technical challenges. She describes the difficulties faced during the Robonet project, where experiments had to be moved between buildings due to renovations, affecting the clarity of results 3. Additionally, Finn highlights the challenge of underfitting in video prediction models, which struggled to fit even the training data due to the diverse and extensive dataset 4. These setbacks reveal the complexities of conducting large-scale robotic experiments.

Normally you hear like overfitting and massive over parameterization and so forth.

--- Chelsea Finn

Despite these challenges, Finn's work continues to push the boundaries of what's possible in robotics research.