Robot manipulation under uncertainty

Thesis

Robots in unstructured environments must perceive, plan, and act under significant uncertainty. Unstructured environments are characterised by cluttered, dynamic, and partially observable settings, as found in homes, offices, and public spaces. To operate effectively in such environments, robots must frequently interact with unknown objects; making and breaking contact as they push, deform, or reorient them during manipulation. Anticipating the outcome of these contact-rich interactions is critical for robot autonomy but remains challenging due to the discontinuous nature of contact dynamics and the uncertainty arising from unmodelled physical properties of the environment. This thesis addresses the problem of controlling robots to exploit contact as a purposeful means of manipulation. In the first part, we formulate the problem of contact-rich manipulation as a model predictive control problem and show that stochastic optimisation in combination with informative priors and learned proposal distributions enables efficient exploration of the contact space in real-time without relying on local gradients or discretisation of the contact space. The ability to replan at high frequency allows the robot to adapt online to changing dynamics, making it implicitly robust to uncertainty. In the second part of this thesis, we extend this framework to explicitly account for uncertainty in the robot’s state and environment. We introduce chance constraints to ensure probabilistic constraint satisfaction and develop a sampling-based approximation that enables efficient evaluation of these constraints within the model predictive control loop. Finally, we explore belief space control through contact, enabling the robot to actively reduce uncertainty in its environment by seeking out informative contact interactions. This allows the robot to not only react to uncertainty but also to explicitly reason about it and take actions to reduce it. Together, these approaches provide a coherent framework that combines reactivity with principled decision- making under uncertainty, enabling efficient and robust robot control strategies that can actively manage and mitigate uncertainty in real-time. We demonstrate the effectiveness of these methods in dynamic robot handover tasks and contact-rich manipulation with a robot manipulator, as well as a quadruped robot with an arm.

Authors: Brudermüller, L

• DOI: 10.5287/ora-dxx8vezp6
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: trajectory optimisation; robust control; contact-rich manipulation; stochastic dynamics
• Subjects: Robotics; Control