In modern manufacturing, human–robot collaboration combines human creativity and problem-solving capabilities with the precision and repeatability of robots. Robotic kitting is a relevant industrial case: while a human operator performs complex tasks such as engine assembly, a robot assists by collecting and organizing components. However, flexible manufacturing makes this task challenging for conventional pick-and-place programs, as components are stored in cluttered bins on shelves rather than in fixed, precisely known positions. Reliable picking therefore requires advanced perception to estimate component poses and grasp-planning strategies that account for collisions, surrounding objects, container boundaries, and component geometry. Although recent advances in machine learning (ML), computer vision, and motion planning offer promising solutions, robust manipulation in industrial environments remains an open challenge. This project aims to bridge the gap between laboratory-tested methods and their deployment in industrial settings. To this end, we use the [WARA Robotics Challenge 2026](https://wara-robotics.se/wara-robotics-challenges/wara-robotics-challenge-2026/) as a testbed for developing and evaluating robust robotic manipulation methods under realistic industrial conditions. Our contribution is an end-to-end vision-to-action pipeline comprising a UR10e robotic arm, a Robotiq 2F-140 gripper, and a wrist-mounted camera. The pipeline consists of three main components: perception, learning-from-demonstration-based motion planning, and machine-learning-based grasp selection. The perception module estimates object poses from RGB-D data. Using kinesthetic demonstrations collected directly on the robot, we learn a dynamical system (DS) motion planner that generates trajectories while avoiding obstacles, including bins, shelves, and self-collisions, and ensuring stable convergence to the target position. The main technological challenge is robust grasp selection in cluttered environments. Valid grasps depend not only on object geometry but also on neighboring objects, bin boundaries, gripper approach direction, contact conditions, and friction. Consequently, candidate grasps may lead to collisions or fail to retain a component during lifting. To address this challenge, we are implementing state-of-the-art grasp-generation methods that use object poses and RGB-D information to generate grasp candidates. A neural network (NN) will then rank these candidates according to their predicted probability of success. The network will be trained using reinforcement learning within a digital twin developed in Unity and supported by the Algoryx physics engine. Finally, the complete pipeline will be evaluated in simulation and validated on the physical experimental setup provided by the WARA Robotics Challenge 2026.