Model Predictive Control with Efficient Trajectory Optimisation for Contact-based Manipulation

This thesis is concerned with enabling robots to manipulate their environments in efficient ways, similar to how humans do. Humans employ a variety of manipulation actions to efficiently manipulate their environments using non-prehensile manipulation. However, enabling robots to make use of these efficient non-prehensile manipulation actions is challenging, particularly when tasking robots to manipulate cluttered environments. There are a variety of issues that need to be addressed to enable robots to reliably operate in cluttered environments. These include, but are not limited to, difficulty predicting contact-interactions between multiple objects, as well as the curse of dimensionality that occurs from scaling levels of clutter.

General purpose physics simulators can be used to predict, to differing levels of accuracy, how multiple objects will interact with one another, subject to the motion of a robotic manipulator. Using these physics simulators, trajectories can be computed to enable a robotic manipulator to perform complex manipulations in clutter to achieve some objective. However, these physics predictions can never perfectly emulate the real world. If trajectories that were computed inside a physics simulator are executed on real robotic hardware open-loop, they are often bound to fail.

A general method to address this issue is model predictive control (MPC). MPC can account for the stochastic difference between a physics simulator and the real world by constantly re-planning trajectories from the current state of the real world. The key idea is that this re-planning needs to occur fast to be effective. Current methods for planning trajectories using physics simulators can be computationally costly, especially when considering manipulation in clutter.

This thesis addresses the issue of computational expense for high dimensional manipulation in clutter tasks, specifically focusing on trajectory optimisation techniques. Two main methods are investigated. Firstly, this thesis explores how approximations can be used to accelerate the computation of dynamics derivatives required by certain trajectory optimisation methods. Secondly, this thesis examines how the dimensionality of a trajectory optimisation problem can be adjusted online during task execution, leveraging the fact that not all objects in a scene are relevant at all times. The thesis concludes by investigating the challenges involved in transferring solutions from simulation to real-world execution, demonstrated through a packing through clutter task.