PAC-Bayes Machine Learning Approach Gives Robots a Fighting Chance in Unfamiliar Surroundings

Probably Approximately Correct (PAC)-Bayes framework lets robots generalize machine learning algorithms for new environments.

Researchers from Princeton University have demonstrated a novel approach to machine learning, which gives robots an edge in safety and success rate when placed into unfamiliar settings: PAC-Bayes.

"Over the last decade or so, there’s been a tremendous amount of excitement and progress around machine learning in the context of robotics, primarily because it allows you to handle rich sensory inputs," says Assistant Professor Anirudha Majumdar of the project. The problem: Overfitting of training data which makes the algorithms perfectly suited to the exact conditions they have been trained with, but ill-suited to novel scenarios."

"The key technical idea behind our approach is to leverage tools from generalization theory in machine learning by exploiting a precise analogy (which we present in the form of a reduction) between generalization of control policies to novel environments and generalization of hypotheses in the supervised learning setting," the researchers explain of the new approach which aims to solve this problem. "In particular, we utilize the Probably Approximately Correct (PAC)-Bayes framework, which allows us to obtain upper bounds that hold with high probability on the expected cost of (stochastic) control policies across novel environments."

The PAC-Bayes system offered good results in testing, but there's room for improvement. (📹: Majumdar et al)

The PAC-Bayes approach was tested in a variety of robotics scenarios, including a wheeled robot, a robotic arm, and a Parrot Swing drone — the latter navigating a 60-foot-long corridor littered with cardboard obstacles. During this test, the guaranteed success rate came out at 88.4 percent with obstacle avoidance in 18 of the 20 trial runs.

"There are a lot of special assumptions you need to satisfy, and one of them is saying how similar the environments you’re expecting to see are to the environments your policy was trained on," co-author Alec Farid notes of the key issue in applying machine learning from other areas to robotics. "In addition to showing that we can do this in the robotic setting, we also focused on trying to expand the types of environments that we could provide a guarantee for."

"The kinds of guarantees we’re able to give range from about 80% to 95% success rates on new environments, depending on the specific task, but if you’re deploying [an unmanned aerial vehicle] in a real environment, then 95% probably isn’t good enough. I see that as one of the biggest challenges, and one that we are actively working on."

The paper has been published under open-access terms on arXiv.org; Princeton University also has details on two other related papers for imitation learning and vision-based planners in robotics.

Gareth Halfacree
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.
Related articles
Sponsored articles
Related articles