The capability of robots to rapidly adapt to new tasks without extensive reprogramming offers significant flexibility in reconfiguration of manufacturing processes to cope with unforeseen events. In modern manufacturing environments where numerous hardware and software systems exchange data with each other to perform a myriad of tasks, modularizing sub-systems and reusing commonly available information like product CAD models can increase robustness and efficiency of the reconfiguration. Yet, current approaches for robotic grasping tend to focus on standalone vision-based learning that often require either retraining to adapt to new object categories or massive dataset not available in manufacturing environments, making generalization challenging. This paper addresses the problem of exploiting available information, like CAD models, in manufacturing settings to efficiently generate a tractable set of grasps for known rigid objects, which can be directly applied to a wide class of robotic manipulations. In order to quickly produce diverse grasp configurations for arbitrary geometric models, we present GraspMixer, a combination of (1) an efficient offline sampler that utilizes specifications of a parallel-jaw gripper, and (2) a mapping function that fuses multiple features of a grasp to output a binary quality metric. During evaluation using physics-based simulations, a robotic gripper successfully executes 92.9% of all grasp configurations for 12 novel objects selected by GraspMixer. Among five different grasp sampling methods, GraspMixer also achieves the highest grasp success rate when performing table-top single object grasping under object pose uncertainty. The computation of this offline pipeline takes less than 1.0 minutes for each object without GPU hardware acceleration, which is comparable to or outperforms most of the benchmarks in the evaluation. Importantly, our framework exhibits impressive simulation-to-reality adaptation, achieving over 95% grasp success rate on previously unseen novel objects. All of these results are achieved with fewer than 10% of the samples typically used by other learning-based grasping techniques. Note to Practitioners—Modularization is a major theme in current manufacturing systems to increase efficiency. In this work, we introduce a new framework called GraspMixer, which is part of a larger manipulation and decision making architecture to enable versatile robotic manipulation in a manufacturing environment. The framework decomposes the task of reasoning about graspable local surfaces on object 3D models into sequentially connected sub-components. GraspMixer leverages information about objects and grippers, including their 3D models, materials, and inertial properties, which are available in a manufacturing environment. This enables our framework to automatically precompute grasping points on new objects that can be shared among multiple robots equipped with parallel-jaw grippers. GraspMixer synergizes with Internet of Things (IoT) and Cloud Computing platforms to efficiently scale up advanced robotic automation in manufacturing. Such a combination could provide greater flexibility in deploying advanced perception systems in a manufacturing environment to accelerate adaptation of the automation while saving computational resources of onboard processors within robots.

Robotic manipulation of deformable linear objects (DLOs) is an active area of research, though emerging applications, like automotive wire harness installation, introduce constraints that have not been considered in prior work. Confined workspaces and limited visibility complicate prior assumptions of multi-robot manipulation and direct measurement of DLO configuration (state). This work focuses on single-arm manipulation of stiff DLOs (StDLOs) connected to form a DLO network (DLON), for which the measurements (output) are the endpoint poses of the DLON, which are subject to unknown dynamics during manipulation. To demonstrate feasibility of output-based control without state estimation, direct input-output dynamics are shown to exist by training neural network models on simulated trajectories. Output dynamics are then approximated with polynomials and found to contain well-known rigid body dynamics terms. A composite model consisting of a rigid body model and an online data-driven residual is developed, which predicts output dynamics more accurately than either model alone, and without prior experience with the system. An adaptive model predictive controller is developed with the composite model for DLON manipulation, which completes DLON installation tasks, both in simulation and with a physical automotive wire harness.