Leveraging Mechanics for Multi-step Robotic Manipulation Planning

Abstract

This thesis focuses on enabling robots to robustly perform complex, multi-step manipulation tasks, like chopping vegetables or wielding a wrench. Completing such tasks requires a robot to plan and execute long sequences of actions, where each action involves many connected, discrete and continuous choices that are critically impacted by constraints relating to force, motion and contact. To tackle this, this thesis contributes models and algorithms that exploit the physics and geometry of the world in order to address the dual challenges of long-horizon decision-making and acting under uncertainty. We apply this in the context of three domains: in-hand manipulation, forceful manipulation and briefly-dynamic manipulation.

First, to reorient a grasped object, we develop a sampling-based motion planner to generate sequences of pushes that slide the object in-hand. We derive an abstraction for pushing to enable the planner to reason about frictional constraints. Second, we focus on forceful manipulation tasks, such as opening a childproof medicine bottle or twisting a nut on a bolt, where the robot's planning choices are impacted by the need to exert force. We define constraints that explicitly consider torque and frictional limits and integrate these into an existing task and motion planning framework. We leverage cost-sensitive planning to enable the robot to generate plans that are robust to uncertainty in the physical parameters. Finally, we frame planning with dynamic actions, like shoveling or toppling, as requiring the robot to reason about both action uncertainty and potential dead ends. We learn a simple action model and formulate a sample-based manipulation planner that guards against dead ends in the face of uncertainty. Throughout this thesis, we validate the practical applicability of our model-based approaches by evaluating them on real robots.