Robotic assembly systems are widely implemented across various industries due to their potential to enhance productivity, precision, and automation efficiency. In particular, a dual-manipulator system offers significant advantages with its wide worksp...
Robotic assembly systems are widely implemented across various industries due to their potential to enhance productivity, precision, and automation efficiency. In particular, a dual-manipulator system offers significant advantages with its wide workspace and high degrees of freedom, enabling dexterous manipulation capabilities. Recently, the robotic assembly has expanded from structured environments to unstructured environments where uncertainties are present in the workspace or workpieces. These uncertainties primarily cause positional misalignment or unknown contact conditions between handling objects, which can lead to assembly failures. Therefore, robust search strategies and contact state estimation methods are essential to ensure successful assembly. In that sense, this thesis introduces two key contributions: (i) robust and effective search trajectory generation and (ii) contact state estimation for determining subsequent manipulator actions.
Firstly, the assembly task is represented as a peg-in-hole assembly, which can describe a wide range of assembly problems. Based on this description, this paper suggests two different search strategies: Twisting motion and Bi-spiral. The twisting motion addresses a dual peg-in-hole assembly where both positional and orientational alignment are required. To achieve this, the motion consists of a planar movement following a spiral trajectory and a rotational movement for alignment with the dual hole. This search strategy provides a distinct advantage as it simultaneously mitigates both translational and rotational alignment uncertainties. The bi-spiral search strategy is a method where both manipulators actively participate in the search task by generating the spiral trajectories. In contrast to the previous research, which either relied on a single arm while fixing the counterpart or lacked a rigorous geometric analysis of relative motions, the bi-spiral trajectory can leverage the dexterous capability of the dual manipulator system.
Secondly, two different contact state estimation methods are presented: an estimation method using a rule-based Gaussian Mixture Model (r-GMM) and an anomaly detection-based estimation (CC-STAR). Since the peg-in-hole assembly consists of sequentially ordered sub-tasks (e.g., approach, search, and insertion), the manipulator should determine appropriate actions based on the current contact state. In this context, the GMM-based estimation method accounts for the non-stationary behavior and correlation among the measured input data to achieve accurate contact state estimation. However, such a data-driven method heavily relies on balanced datasets, which are difficult to obtain due to the short duration of certain contact states and the rarity of failure cases. To address this challenge, CC-STAR employs an anomaly detection technique with domain knowledge to identify various contact state transitions.
Finally, all proposed methods are validated through the real-world robot experiment where the assembly environment consists of the torque-controlled manipulators and commercial prefabricated furniture.