Towards an holistic approach for highly flexible robotic assembly systems

Robson, Mark, Andrew (2024). Towards an holistic approach for highly flexible robotic assembly systems. University of Birmingham. M.Res.

Preview

Robson2024MRes.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (17MB) | Preview

Abstract

This thesis presents a general framework for the control of a robotic agent tasked with solving problems in the domain of industrial assembly. A set of motivating example problems are described which shape the development of the framework. The framework fuses the concepts of a closely coupled, multi-layer reasoning architecture with abstract assembly sequence planning. This framework provides a reusable and flexible structure for robotic manipulation via a modular location graph, and a grasp planning mechanism that considers object and task constraints.

One example assembly task inspired by the building of light gauge steel frame panels for the construction sector is used to develop a benchmark object set and challenge assemblies. These objects are published as the open source Robotic Assembly Manipulation and Planning (RAMP) benchmark [1]. The abstract task planning layer finds valid sequences of part addition to complete an assembly task given a task description specified using Action Language ALd [2] sequentially feeding these steps as subgoals to a closely coupled robotic reasoning framework based on the REBA architecture [3]. Experiments demonstrate application to long time horizon assembly problems requiring the addition of many parts with over 200 low level robot actions. The abstract task planning layer reduced coarse-resolution planning time by 93.5% compared to a baseline which must simultaneously consider assembly sequencing and robot actions. The modular location graph structure links positional information between the task space and logical domain. We utilise this structure to parameterise the assembly actions of the RAMP benchmark tasks. Additionally, a pruning heuristic is proposed to speed up searching in this location graph when the robot is planning assembly motions.

A grasp planning approach is detailed utilising a weighted grasp scoring model considering a combination of measurement uncertainty and variation in the extracted surface, the contact angle of gripper fingers to the surface, as well as task constraints. A three-level representation for objects, compatible with our framework, includes object class membership, point cloud data representing the objects surface, and semantic keypoints linked to the object parts. A learned model is used to encode task-specific knowledge from a small number of exemplars of objects, tasks, and relevant grasps; preserving the relationship between keypoints and grasp points for specific tasks despite changes in factors such as the scale and orientation of objects. The learned models are queried at run time to guide the generation of grasps that balance task and stability constraints. Through experimental evaluation on a Franka robot manipulator with a parallel gripper, it is demonstrated that this method is able to generate grasps on previously unseen objects achieving the desired task-specific trade off whilst maintaining a high degree of grasp stability.

Type of Work:

Thesis (Masters by Research > M.Res.)

Award Type:

Masters by Research > M.Res.

Supervisor(s):