Robust physics-based robotic manipulation in real-time

This thesis presents planners and controllers for robust physics-based manipulation in real-time. By physics-based manipulation, I refer to manipulation tasks where a physics model is required to predict the consequences of robot actions, for example, when a robot pushes obstacles aside in a fridge to retrieve an object behind them.

There are two major problems with physics-based planning using traditional techniques. First, uncertainty, both in physics predictions and in state estimation, can result in the failure of many physics-based plans when executed in the real-world. Second, the computational expense of making physics-based predictions can make planning slow and can be a major bottleneck for real-time control. I address both of these problems in this thesis.

To address uncertainty, first, I present an online re-planning algorithm based on trajectory optimization. It reacts, in real-time, to changes in physics predictions to successfully complete a manipulation task. Second, some open-loop physics-based plans succeed in the real-world under uncertainty. How can one generate such robust open-loop plans with guarantees? I provide conditions for robustness in the real-world based on contraction theory. I also present a robust planner and a controller. It autonomously switches between robust open-loop execution, and closed-loop control to complete a manipulation task. Third, a robot can be optimistic in the face of uncertainty. It can adapt its actions to the accuracy requirements of a task. I present such a task-adaptive planner that embraces uncertainty, pushing fast for easy tasks, and slow for more difficult tasks.

To address the problem of computationally expensive physics-based predictions, I present learned and analytical coarse physics models for single and multi-object manipulation. They are cheap to compute but can be inaccurate. On the other hand, fine physics models provide the best predictions but are computationally expensive. I present algorithms that combine coarse and fine physics models through parallel-in-time integration. The result is orders of magnitude reduction in physics-based planning and control time.