| dc.description.abstract | In this thesis, we address the problem of long-horizon robotic manipulation under partial observability. Tasks such as gearbox assembly or tidying a workstation involve many objects and necessitate a variety of manipulation capabilities. These long-horizon tasks are commonly addressed by hierarchical approaches, which introduce state and action abstractions to make planning tractable. However, our abstractions often rely on imperfect models of the world, which can lead to brittle execution. Furthermore, these abstractions depend on having accurate state information, which is often only noisily sensed, if sensed at all. For example, in the assembly domain, the pose of each part may only be known within a few millimeters, and a box’s mass distribution may be completely unsensed. To deploy robots outside of structured environments like the factory, they will need to be robust to model misspecification and partial observability. The central idea of this thesis is that we can develop adaptive abstractions to improve the robustness of hierarchical planning once the robot is deployed. Adaptive abstractions incorporate observations from the real world that are informative about misspecifications and partial observability, essentially allowing the planner to adapt to its deployment environment. We explore this idea by developing three types of models that enable this adaptivity at different levels of the abstraction hierarchy: plan feasibility models, adaptive samplers, and reactive control policies. In our first contribution, we consider adding adaptivity to a task and motion planning system at the task-planning level. We focus on a setup where the robot has access to a set of parameterized skills, but these skills are derived from imperfect models. To enable robust planning, we propose to autonomously learn skill feasibility models once the robot is deployed through a curious exploration phase. Critically, we propose a novel active learning framework to enable efficient learning without human intervention. We show that the resulting feasibility model leads to robust task performance on multiple downstream tasks in a stacking domain. Our second contribution looks at developing adaptive samplers that can incorporate information about object state that is typically unobserved (e.g., inertial and frictional properties). General-purpose belief representations can handle this partial observability, but online inference is computationally expensive. Instead, we propose to use an offline phase to learn an inference network that directly predicts a distribution over object properties that is consistent with the interaction history. We show that inference networks enable efficient adaptation in a grasping domain with heavy objects. Our final contribution focuses on learning adaptive controllers such that robustness is handled at the lowest level of the abstraction. We consider precise contact-rich manipulation tasks that are sensitive to pose estimation errors. To overcome noisy poses at the control level, explorative contact is necessary, but unintended forces can lead to catastrophic outcomes such as part slippage or damage. We propose to use simulation in an offline phase to train reactive force-aware policies. The policies are trained to overcome pose uncertainty while using force-sensing to adaptively limit excessive forces. The result is robust real-world performance on the multistage assembly of a planetary gearbox system, which includes insertion, gear-meshing, and nut-threading tasks. In summary, adaptive abstractions can be used to increase the robustness of hierarchical manipulation planning, an important step in deploying robots outside of the lab or factory. Throughout the thesis, we validate the proposed approaches on the real robot in stacking and assembly domains. | |
