Overcoming Epistemic Uncertainty to Plan with Learned Dynamics Models for Robotic Manipulation

Manipulation of objects like cables, cloth, and scattered rigid objects is essential in a broad range of settings, including factories, homes, and hospitals. While modern robots are physically capable of manipulating these objects, they lack the algorithms necessary to understand how these objects move when being manipulated. To give robots the ability to perform a wide range of manipulation tasks necessary for maintenance, manufacturing, and cleaning tasks, this project will develop new methods that allow robots to learn how these objects move from data.

Yet there will always be some uncertainty in what the robot learns, and if the robot is over-confident in its understanding, it may make many errors, or even be unable to complete the task-at-hand. Thus a robot requires methods to reason about what it has and has not learned so that it can accomplish useful tasks reliably. Finally, as the robot performs the manipulation task-at-hand it will acquire experience of manipulating that particular object.

The robot will need to use that experience to enhance its understanding of how the object moves, leading to more reliable performance. Endowing robots with the ability to learn from data while being aware of the uncertainty in what they learn and improving their understanding from experience will enable a wide range of robotics applications across many sectors of industry.

To provide robots with these fundamental capabilities, this project will build a much-needed bridge between the fields of dynamics learning and motion planning, enabling roboticists to take advantage of the latest dynamics learning methods to plan for manipulation tasks that are currently considered too difficult to model analytically. Recent advances in machine learning have allowed dynamics models to be learned from high-dimensional data, such as images.

However, these learned models are currently insufficient for planning because they do not account for epistemic uncertainty, i.e. uncertainty due to a lack of data. Not considering epistemic uncertainty leads to unreliable estimates of a model's confidence in its prediction, which can cause highly state-dependent errors. Furthermore, fundamental advances in motion planning are required to robustly plan with models that are not guaranteed to be valid everywhere.

Thus this project will explore foundational methods for 1) estimating the confidence of a dynamics model's prediction while accounting for epistemic uncertainty; 2) improving dynamics predictions using limited data during execution; and 3) principled motion planning that uses these predictions and confidence estimates to avoid areas of the state space where the model is unreliable. The key insight that enables tackling this difficult problem is that a dynamics model need not be globally-accurate to be useful for planning motion.

The effectiveness of these methods will be demonstrated by integrating them into a framework that allows robots to manipulate objects such as rope, cloth, and debris for a wide range of practical tasks.

This project is supported by the cross-directorate Foundational Research in Robotics program, jointly managed and funded by the Directorates for Engineering (ENG) and Computer and Information Science and Engineering (CISE).

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.